Startup Showcase Registry: Innovating Cloud and AI Security
Discover innovative AI and cloud security solutions designed to tackle today's biggest challenges.

Explore a curated registry of cutting-edge solutions from startups addressing critical cloud security threats and advancing AI governance.
Designed for enterprises and security leaders, this resource provides actionable insights to mitigate risks, ensure compliance, and drive innovation across the industry.
Explore Solutions
Enkrypt AI
As enterprises deploy agentic AI systems, autonomous agents that call tools, access data, and chain decisions across workflows, traditional identity and access management frameworks were not designed to govern non-human, AI-driven actors. The CSA's guidance on Agentic AI IAM highlights a critical gap: agents need their own identity, scoped permissions, behavioral accountability, and runtime controls to prevent privilege escalation, prompt injection, and unintended tool execution.
Enkrypt AI addresses this challenge directly. Our AI Security and Compliance Platform operates inline between users, agents, and models, enforcing guardrails on every input, output, and inter-agent communication in real time. Rather than replacing IAM infrastructure, Enkrypt AI extends it into the AI layer: governing what agents can say, do, and access based on enterprise-defined policies.
Key capabilities aligned to the CSA Agentic AI IAM framework include runtime policy enforcement that constrains agent behavior to authorized actions; full tracing of agent decisions, tool calls, and hand-offs for auditability; kill-switch and sandboxing controls to halt unsafe autonomous execution; and supply chain risk assessment across Model Control Points (MCPs), where Enkrypt AI has scanned over 12,000 MCP servers globally to detect exposed agent interfaces.
The platform supports SaaS, VPC, and on-premises deployments, ensuring agent governance can be enforced regardless of where AI workloads run. All controls integrate with existing SIEM, GRC, and IAM ecosystems, and compliance outputs are mapped to NIST AI RMF, OWASP LLM Top 10, MITRE ATLAS, and the EU AI Act.
Enkrypt AI enables enterprises to deploy agentic AI with confidence, not by slowing agents down, but by ensuring every action they take is authorized, observable, and compliant.
Enkrypt AI
Enkrypt AI secures enterprise AI with runtime guardrails, red teaming, and compliance automation across text, image, voice, and agentic systems. SOC 2 Type II and ISO 27001 certified.
Learn morePowerQuant ApS
Keelstamp is software built and owned by PowerQuant ApS, a bootstrapped Danish company founded in February 2026. It sits between the AI automations that agencies and MSPs run for their clients (n8n, Make, Zapier, GoHighLevel) and the real world, and gives the end client, the principal, control over what the agent does on its behalf.
How it relates to CSA's Agentic AI work:
1. Oversight by the principal. Each end client gets its own approval inbox and stop button under the agency's brand, by magic link on mobile, with no n8n login. Actions that need a human decision wait for it, and the client can stop the automation at any time.
2. Hard budget caps that fail closed. A monthly cost cap per end client is checked before every model call. If the check cannot be made, the call is refused.
3. Verifiable execution records. Every AI action produces a signed receipt (COSE_Sign1, Ed25519, RFC 9943). The client verifies receipts offline with our open-source verifier (github.com/keelstamphq/verifier, Apache-2.0) without trusting the operator. Daily signed checkpoints are published in a public repository (github.com/keelstamphq/transparency), so a receipt cannot be altered or removed later without detection.
4. Data handling. Core data runs in the EU (Hetzner, Germany; Supabase, Frankfurt). The monthly report is deterministic: no LLM reads client data. Our subprocessor list is public.
Receipts and logs can serve as documentation of human oversight and log retention where the obligation sits with the client. Keelstamp does not classify the client's system and makes no compliance claims.
Status: v1 is not yet released. A Founding Partner pre-sale is open, n8n templates are published, and the verifier and transparency log are public on GitHub. Three co-founders, no external funding. PowerQuant ApS is an EU AI Pact signatory and is listed by the European Commission as a signatory of the Code of Practice on transparency of AI-generated content; neither is a certification.
PowerQuant ApS
Keelstamp: approval inbox, hard cost caps and offline-verifiable signed receipts for the AI automations agencies run for their clients.
Learn moreUltra Security
As AI agents take on real work across organizations, they're connecting to tools, data, and systems through MCP at a scale and speed that no existing security infrastructure was built to handle. Every connection is a potential attack surface. Every tool call is a risk. Most organizations have no visibility into any of it.
Ultra is the security layer for AI and MCP. Deployed as a locally hosted gateway and firewall, Ultra sits between every AI client and MCP server in an environment, giving security teams real-time visibility, guardrails, and audit logging across every agent action, tool call, and data access, from a single platform. It installs or is deployed via MDM in seconds and connects to any AI client or MCP server through a single universal proxy.
Ultra's Guardrails enforce dynamic, policy-driven controls at the proxy layer, including data loss prevention, tool allowlisting, and custom rules that adapt to the environment. Anomaly Detection surfaces behavioral deviations the moment they happen, not after the damage is done. Ledger gives security teams a continuously updated health and risk picture of every MCP server in the environment, tracking posture, activity, and trust signals across the entire connected surface. And Hub provides centralized visibility and RBAC across every MCP server and AI client an organization operates. Ultra is SOC 2, ISO 27001, ISO 42001, and NIST AI RMF certified from day one.
Agents are already working. Ultra makes sure teams know what they're doing.
Ultra Security
The security layer for AI and MCP. Ultra gives teams real-time guardrails, visibility, and audit logging across every agent action and tool call.
Learn moreCyberVisir Solutions Ltd
Security teams using AI still need verifiable evidence of exposed accounts and data. LeakData, developed by CyberVisir Solutions Ltd, brings breach-exposure and credential-risk investigation into ChatGPT and existing security workflows through its web platform and API.
Our ChatGPT app is listed in the Security category. A connected user can check the connected account's verified primary email, review breach sources, dates and affected data types, and work through a follow-up action plan. Public breach-catalog lookups are also available. The app does not expose raw leaked records or search another person's private breach data. A negative result is not proof that an account is secure.
Our closest alignment is CSA's Data Security within AI Environments: understanding data exposure and handling security evidence carefully in AI-assisted work. We selected Observability to describe visibility into external exposure. Our scope is not agent-runtime telemetry, agent identity issuance, authorization enforcement, prompt-injection prevention or DLP.
We are a two-person startup. Our co-founder Emre Çapan's published FreeBSD Journal article, "Auditable Credential-Exposure Checks on FreeBSD Without Sending Full Passwords," examines privacy-conscious credential checks and auditable evidence. We welcome CSA's review of LeakData as a complementary exposure-intelligence tool for AI-assisted security investigations.
ChatGPT listing: https://chatgpt.com/plugins/plugin_asdk_app_6a679a2771648191a133ec7977c547ec
Article: https://freebsdfoundation.org/our-work/journal/browser-based-edition/production-deployments/auditable-credential-exposure-checks-on-freebsd-without-sending-full-passwords/
CyberVisir Solutions Ltd
Breach-exposure and credential-risk intelligence for AI-assisted security investigations through ChatGPT, web and API.
Learn moreAgentic Control Plane
CSA's research shows only 26% of organizations have comprehensive AI governance. Yet nearly 60% plan to deploy agentic AI within 12 months. The gap between adoption and governance leaves enterprises exposed.
As CSA's Agentic AI IAM paper identifies, traditional protocols like OAuth 2.1 and SAML were not designed for autonomous agents: they're coarse-grained, static, and create "confused deputy" vulnerabilities where agents inherit broad permissions with no clear accountability chain. AI workloads already initiate 148x more authentication requests per hour than humans. Existing IAM infrastructure wasn't built for this.
Agentic Control Plane (ACP) addresses the three-party identity problem at the core of agentic AI security: when a user authorizes an agent to act on enterprise systems, the receiving system has no way to verify whether that action was sanctioned, scoped, or still valid. ACP provides verified agent identity separate from the user, full delegation chain tracking and least-privilege policy enforcement at runtime, and an enterprise admin console for oversight and revocation. ACP's policy enforcement layer operates independently of the underlying model or agent framework, making governance controls vendor-neutral by design.
ACP directly implements CSA AICM control DSP-17, which explicitly requires monitoring and logging all agentic AI and non-human identity access to sensitive data. It also indirectly addresses DSP-28 (Shadow AI Detection) — as CSA's Data Security guidance notes, banning AI outright drives employees toward shadow IT. ACP gives IT the visibility to sanction agentic adoption confidently.
ACP's open-source foundation (GatewayStack, published on npm) includes active contributions to W3C WebMCP standards discussions on agent identity.
Agentic Control Plane
Agentic Control Plane (ACP) governs agentic AI through verified agent identity, delegation chain tracking, and runtime policy enforcement — making autonomous agents auditable and controllable.
Learn moredwaar.ai (DwaarAI Tech Pvt. Ltd.)
Dwaar helps enterprises discover, test and govern AI agents across employee endpoints, SaaS platforms and the applications they build. It connects security policies to the tools, data and actions an agent can use.
Aligned with CSA's The State of AI Security and Governance, Dwaar makes governance operational through an inventory of supported agents and their capabilities, organisation-specific policies, and evidence for investigating policy violations. Security teams can review available ownership and access metadata, identify unapproved connections, and apply controls at supported integration points.
dForge assesses agentic workflows for prompt injection, tool misuse, data exfiltration and excessive agency. Application context and findings inform targeted policies. dFence evaluates requests and tool actions against the agent's intended purpose and configured data boundaries, then allows, blocks, redacts or alerts according to policy. Deterministic and model-based guardrails support controls across supported environments.
This approach also addresses themes in CSA's Agentic AI Identity and Access Management and Data Security within AI Environments: accountable ownership, context-aware authorisation, protection of sensitive data and traceable actions. For example, a support agent may need customer records for a billing query; a configured policy can prevent an injected instruction from turning that access into an unrelated bulk export.
Dwaar brings adversarial testing, runtime policy enforcement and audit evidence into a shared workflow, helping security teams assess risk before deployment and govern agent behaviour in operation. Coverage and enforcement depend on the connected platform and configured control points.
dwaar.ai (DwaarAI Tech Pvt. Ltd.)
Dwaar discovers, red-teams and governs AI agents with runtime guardrails for tool misuse, prompt injection and sensitive data exposure.
Learn moreQlok Inc
QATI is a pre-execution governance and trust-validation layer for enterprise AI workflows.
As AI agents, copilots, automation, and workflow systems move from generating outputs to taking actions, enterprises need controls that determine whether an AI-driven action should proceed before it reaches tools, APIs, data, approvals, infrastructure, security operations, financial systems, or production environments.
QATI evaluates the full action path before execution, not just the prompt or final model output. The platform analyzes signals across identity, session state, workflow context, retrieval quality, model behavior, tool/API authorization, approval status, policy state, anomaly patterns, and audit integrity.
In structured validation testing, QATI distinguished normal workflow activity from rising risk, anomaly spikes, sustained trust degradation, and unsafe execution conditions before actions reached production. It identified conditions including context drift, prompt-injection pressure, unauthorized tool/API use, low-trust retrieval, missing approvals, and workflow-risk escalation.
QATI SDK v1.0.1 has been published with expanded signal support and advisory responses. The system is designed to help enterprises determine whether an AI-driven action should proceed, escalate, require approval, be logged, or be blocked before execution impact.
QATI is vendor-neutral and designed to support enterprise AI governance, agentic AI security, auditability, and pre-execution control across high-consequence AI workflows.
Qlok Inc
QATI validates AI action paths before execution, detecting risk, drift, unauthorized tool use, missing approvals, and unsafe workflow conditions.
Learn moreAiceberg
The CSA's State of AI Security and Governance survey and Data Security within AI Environments publication share a common finding: organizations are deploying LLMs and agentic AI at scale without the controls needed to protect data across the AI lifecycle. Aiceberg is purpose-built to close that gap.
As a real-time Guardian Agent trust layer, Aiceberg directly addresses four of the six control gaps proposed in the CSA AICM framework:
* DSP-25 (Prompt Injection Defense) via input validation, context isolation, and output filtering
* DSP-26 (Model Inversion Protection) through response truncation and entropy monitoring
* DSP-28 (Shadow AI Detection) via SIEM integration that surfaces unauthorized AI tool usage enterprise-wide
* Secure Runtime Environment through least-privilege policy enforcement and deterministic agent guardrails
For agentic AI deployments, Aiceberg continuously monitors business intent signals and agentic instructions to detect behavioral drift and unauthorized actions before harm occurs, providing the containment strategies the CSA identifies as essential for autonomous systems. Aiceberg deploys as a trust layer with flexible API and SDK integrations, fitting into existing development workflows and security stacks including SIEM, CASB, and LLM frameworks with minimal disruption.
Underpinning everything are deterministic, non-generative ML models that make every AI decision traceable, explainable, and audit-ready, directly supporting GDPR, HIPAA, CCPA, and EU AI Act compliance requirements.
Aiceberg
AI security, governance, and observability platform protecting LLMs and agentic AI from threats, data leakage, and shadow AI via a Guardian Agent trust layer.
Learn moreSweet Security
Sweet Security blocks harmful behavior across cloud and AI the moment it happens. The security problem is no longer identifying risk, it is stopping that behavior before it becomes a business-impacting incident. That requires proactive runtime enforcement, not another alert waiting for someone to respond.
This addresses a central challenge in CSA's threat analysis: credentials can confirm that an identity or AI agent is permitted to enter an environment, but cannot prove that its behavior is legitimate. An attacker, compromised workload or manipulated agent can hold valid credentials and still take actions its creator never intended. Sweet evaluates what applications, workloads, identities and AI agents actually do at runtime, where intent becomes action, and enforces in real time.
For AI agents, Sweet can block prompt-injection attempts, terminate sessions on unauthorized tool calls, stop secrets and PII leaving, reduce excessive permissions and quarantine suspicious tool execution. Across the cloud those agents run on, Sweet can revoke abused credentials mid-attack, isolate workloads attempting lateral movement and block command-and-control traffic at the workload.
This cloud-to-AI context is critical. An AI-specific layer may govern prompts and models without seeing what happens when an agent uses a cloud identity, reaches sensitive data or runs a tool in production. Sweet protects the whole path, from prompts, models and agent tools to workloads, identities, APIs and data, in one runtime context.
Sweet's Learning Loop continuously tests the environment, fixes proven risk and strengthens defenses from real application and agent behavior, so protection sharpens as threats evolve.
Sweet secures an organization in two hours, processes more than one billion runtime events per day and enforces inline controls in under 100 milliseconds with over 99% accuracy. The result: every application and AI agent operates exactly as its creator intended, and nothing more.
Sweet Security
Proactive runtime enforcement for cloud and AI. Sweet blocks a rogue agent the moment it acts and turns every fix into a real-time guardrail.
Learn moreEve Security
Eve Security is the runtime security and governance platform for Agentic AI. As organizations deploy AI agents across platforms like Microsoft Copilot, Claude, Cursor, Glean, Salesforce, ServiceNow, Snowflake, Jira, and custom AI applications, they need visibility and control over what those agents actually do, not just who can access them.
Eve provides real-time visibility into agentic activity, discovering AI agents, tools, MCP servers, integrations, and autonomous workflows across the enterprise, including shadow AI. Security teams gain a complete inventory of their AI ecosystem and can understand which agents are acting, which systems they access, what data they use, and the actions they perform.
At runtime, Eve continuously evaluates AI behavior against organizational security, compliance, and business policies. Using intent-aware policy enforcement, Eve can allow, block, modify, or require human approval before high risk actions are completed. This prevents data leakage, unauthorized transactions, excessive permissions, prompt injection attacks, policy violations, and unsafe autonomous behavior before business impact occurs.
Eve secures AI interactions across enterprise applications, SaaS platforms, APIs, MCP servers, coding assistants, and custom AI agents without requiring changes to the AI models themselves. The platform helps organizations safely adopt Agentic AI while meeting governance and regulatory requirements through comprehensive audit trails, policy enforcement, and runtime protection.
Unlike traditional security tools that focus on identity, network traffic, or data posture, Eve protects AI decision making and actions as they happen, providing organizations with the confidence to scale Agentic AI securely.
Eve Security
Gain complete visibility into your agentic AI landscape and secure every agentic activity with real-time governance and control.
Learn moreSilmaril
Silmaril is the first self-healing firewall for AI applications, blocking threats existing defenses are engineered to miss. As a runtime security solution, Silmaril complements governance to deliver comprehensive risk reduction. In just two months, we've prevented 8 verified breaches and blocked twice as many attacks as existing solutions with a p90 latency of 15 ms.
Our firewall continuously improves using real-time, application-specific threat data including inputs, application states, and execution traces. This adaptive approach maintains 99% accuracy against evolving threats, reducing MTTR from days to 1 hour.
Silmaril
Prompt injection defense that self-improves. Block 2x more threats at a 10x lower latency with Silmaril's AI firewall.
Learn morepraedictio.ai, Inc.
Agentic AI is only safe if the software it writes, runs, and reaches can be trusted to do only what it was built to do. As agents scale, they find and chain unknown vulnerabilities at machine speed. The Hugging Face and JFrog Artifactory escape is the clearest example, where autonomous agents exploited unknown zero-days in software to break out. The State of AI Security and Governance concludes that traditional controls are insufficient for the behavioral, non-deterministic nature of AI systems. praedictio.ai answers that with a deterministic control.
Most solutions govern what the agent is allowed to do. We secure the software the agent runs and attacks. Identity, data-protection, and governance controls secure the AI and its data; we secure the software layer they all run on and that agents attack. Program Behavior Intelligence builds a deterministic, auditable behavior schema, a precise specification of exactly what a program is supposed to do. It is not a trained AI model and carries no probability. It is generated frictionlessly from the QA and acceptance tests a team already runs, with no policies to author and no source code required, and it is enforced at runtime so anything outside the schema is prevented before it executes.
This is enforcement, not detection. A crossing is deterministic, a real attack, a real defect, or a gap the tests never exercised, never a probabilistic alert. The same schema secures the software supply chain: when a dependency updates or an AI-generated component ships, we show and enforce what it now does.
We demonstrate this live on a deserialization remote-code-execution flaw in Apache ActiveMQ, prevented with no prior knowledge of the vulnerability, the same class of flaw at the executing end of the Artifactory incident. praedictio.ai is the layer that ensures software behaves as intended, so autonomous agents can operate without being able to turn the software they use against the enterprise.
praedictio.ai, Inc.
An auditable deterministic behavior schema, built from your existing QA tests, that enforces what software may do and prevents anything outside it. Not an LLM
Learn moreInviGrid
InviGrid operationalizes the Cloud Security Alliance’s vision for securing AI by delivering a full-stack, automated control plane for AI infrastructure and execution. It unifies governance, policy definition, secure provisioning, continuous monitoring, and automated remediation in real time.
The CSA document highlights critical gaps in AI security, including Shadow AI detection (DSP-28), prompt injection defense (DSP-25), and the need for runtime controls across the AI lifecycle. InviGrid addresses these by embedding intelligence directly into infrastructure, where environments are governed, created, secured, and operated automatically by InviGrid’s agents.
Instead of requiring teams to manually configure infrastructure and layer on controls, InviGrid allows users to simply request what they need (“I need an environment”), and a secure-by-design, compliant environment is provisioned. It continuously monitors for drift, exposure, and risk, automatically remediating issues to maintain a secure state, aligning with CSA’s emphasis on lifecycle-aware security and continuous enforcement.
InviGrid directly addresses Shadow AI and uncontrolled execution by centralizing how environments are created and operated. All activity, whether human or agent-driven, occurs within defined guardrails while gaps are continuously identified and resolved.
The platform automates CSA-recommended practices such as least-privilege access, auditability, and proactive response to risk. Through policy-driven automation and agent-based execution, InviGrid enables secure AI adoption without slowing innovation.
InviGrid turns CSA’s framework from guidance into execution, ensuring every environment is created secure, stays secure, and auto-remediates. This shifts organizations from chasing risk to eliminating it, enabling fast, scalable AI adoption without sacrificing control.
InviGrid
InviGrid defines infrastructure for the agentic AI era where agents create, secure, and operate governed environments automatically at machine speed and scale.
Learn moreDaylight Security
Daylight is a security services company delivering Managed Agentic Security Services (MASS), including MDR, threat hunting, incident response, and more, through a fundamentally different architecture than traditional security services providers.
Daylight's architecture combines an agentic platform that runs the full cycle from detection to response with security experts from IR and threat hunting backgrounds. The platform collects identity and business context across systems to investigate alerts and continuously learns your environment. Security experts validate decisions, feed insights into the platform, optimize detections, and take over in case of an incident.
The result: security teams move from firefighting to strategic work.
Daylight Security
Daylight MDR combines an AI-native platform with security experts in a unified architecture built for the full SOC cycle. The platform integrates deeply across endpoint, cloud, identity, and SaaS environments, collecting business context to conduct cross-system investigations that reach definitive verdicts, not escalations. Security experts with IR and threat hunting backgrounds engage where human judgment matters: resolving ambiguous cases, customizing detections, building knowledge that improves future investigations, and leading response during confirmed incidents.
Learn moreAsqav
Asqav provides cryptographic governance for AI agents through signed, tamper-evident receipts. Each consequential action and policy decision can be recorded with identity, authorization context, result digest, and independent timestamps. Teams can export audit packs and verify records independently, including offline, without relying on a mutable dashboard.
The product is designed for regulated and high-stakes deployments where security and GRC teams need a reviewable record of what an agent did, what it was authorized to do, and whether the record changed. It complements runtime controls: a policy gate can decide whether an action is allowed, while the receipt preserves the decision and outcome for later review.
Asqav is currently pre-revenue and seeking pilot partners with AI systems in finance, healthcare, enterprise operations, or other high-stakes workflows. Receipts support recordkeeping and independent verification; they do not by themselves establish legal compliance, which depends on the system and jurisdiction.
Asqav
Signed, tamper-evident receipts and offline verification for accountable AI-agent actions and audit packs.
Learn moreAembit
Aembit provides identity and access management for agentic AI and workloads, with control enforced at the moment access is requested rather than through stored credentials and secrets.
Organizations adopting AI agents quickly encounter a structural limitation. Agents require broad access to data, APIs, and internal services, yet existing identity models cannot clearly distinguish between human and agent activity, cannot scope access with precision, and often rely on long-lived credentials placed in application code or configuration. This leads to gaps in attribution, excessive permissions, and a growing credential footprint that is difficult to audit or revoke.
Aembit addresses this by establishing identity as the control layer for agent activity. Each request is evaluated against centrally defined policy using verified workload identity, user context where applicable, and runtime conditions. Instead of distributing secrets, Aembit issues short-lived, scoped credentials for each interaction, eliminating static credentials and reducing the exposure window.
For user-driven agents, Aembit introduces a blended identity model that binds the agent to the user on whose behalf it operates. This allows security teams to attribute every action to both the human and the agent, restoring accountability in environments where agents act autonomously or semi-autonomously.
This model applies consistently across SaaS, cloud, and internal services, including emerging patterns such as Model Context Protocol (MCP) servers, where early implementations often lack strong identity controls, policy enforcement, and secure credential handling .
Aembit integrates with existing identity providers and security tooling, enforcing access through a centralized policy layer. The result is a consistent method for controlling how agents and workloads access sensitive resources, with full auditability and without adding operational burden to development teams.
Aembit
Aembit provides identity and access management for AI agents and workloads, replacing static credentials with policy-enforced, short-lived tokens.
Learn moreNative
AI agents act through cloud identities and permissions. Whatever an agent's role, service account, or managed identity allows, the agent can do, whether it's following instructions or being steered by an attacker. Guardrails inside the model or the prompt are advice a misbehaving pipeline can ignore. The boundary that holds is the one the cloud provider enforces.
Native turns your built-in cloud security controls into operationalized secure architecture across AWS, Azure, Google Cloud, and OCI. Security teams define what must be true once: which services agents can use, which data and environments they can reach, and which actions are structurally unavailable, such as an agent changing its own permissions, exfiltrating data outside the organization, or deploying in unapproved regions. Native maps that intent to the provider-native controls each cloud already enforces, including SCPs, RCPs, organization policies, IAM deny policies, VPC Service Controls, Azure Policy, and OCI compartments.
Because enforcement runs through the cloud's own authorization layer, the limits apply to every identity in the execution path, from the agent to the Lambda, Function, or Cloud Run tool it calls, and hold regardless of which model, framework, or MCP server is involved.
Native manages the full lifecycle: Impact Simulation validates a policy against real historical activity before rollout, Enforcement deploys through guided steps or Terraform and surfaces blocked actions with full call chains, and Exception Management tracks approved deviations. The architecture stays aligned as agents, services, and the organization change.
This directly addresses CSA Top Threats such as misconfiguration and inadequate change control, identity and access management, and insecure third-party resources, and puts the least-privilege, scoped-access model described in CSA's Agentic AI Identity and Access Management guidance into deterministic, provider-enforced practice.
Native
Secure cloud architecture for the AI era: one set of boundaries across AWS, Azure, Google Cloud and OCI that limits what AI agents can reach and do.
Learn morePluto Security
Agentic AI has changed the endpoint. Employees now use AI tools that can call MCP servers, load skills and packages, execute commands, access sensitive systems, and create new agents, apps, and automations. Security teams need to see that full chain and intervene without blocking adoption.
Pluto Security is the AI Workspace Security Platform. Pluto secures the full AI surface on the endpoint: the tools employees run, the ecosystem those tools pull in, and the software they create. It provides agentless discovery, contextual risk analysis, and real-time policy enforcement across all three layers.
Pluto's runtime capabilities analyze agent activity, skills, intent signals, and attack paths in context. Security teams can stop or redirect a risky action, restrict a malicious skill or MCP connection, or prevent access to sensitive credentials without shutting down the entire workflow. Pluto also covers common endpoint blind spots such as IDE extensions, browser plugins, packages, and AI app builders.
This maps directly to CSA's Data Security within AI Environments guidance, particularly its focus on shadow AI detection, third-party components, cross-boundary data access, and secure runtime controls. Pluto uses the endpoint as the control point because that is where the user, AI tool, local and remote components, files, credentials, commands, and resulting actions come together.
Pluto works through the security stack enterprises already operate, including EDR/XDR, identity, MDM, browser, and network tools. The result is enforcement and enablement that let security teams say yes to agentic AI without losing control.
Pluto Security
Security for the AI workspace: discovery, contextual risk, and real-time enforcement across every tool employees build with
Learn moreThirdKey AI
The ThirdKey Trust Stack is an open-source framework that lets enterprises grant AI agents real autonomy without losing control. Its premise is structural: dangerous agent actions should be architecturally inexpressible, not merely detected after the fact.
At the core is Symbiont, a Rust-native runtime built around the ORGA reasoning loop (Observe → Reason → Gate → Act). Phase transitions are typestate-enforced at compile time, so it is impossible to construct a code path from Reason to Act that bypasses the Gate. The Gate itself runs Cedar policies with deny-by-default authorization, evaluating every proposed action against organizational rules outside any LLM influence — Cedar receives only structured inputs, so prompt injection, social engineering, and creative reasoning cannot reach the policy decision. Cedar's five decisions (Allow, Deny, Modify, Step-Up, Defer) cover human-in-the-loop escalation for high-risk operations.
Mapped to CSA thought leadership:
Securing LLM-Backed Systems: Essential Authorization Practices. Cedar at the ORGA Gate is a formally analyzable, deny-by-default authorization engine. SchemaPin cryptographically verifies MCP tool schemas (ECDSA P-256, TOFU pinning) before any tool is callable, closing the supply-chain gap. AgentPin provides domain-anchored ES256 agent identity, so the policy engine evaluates verified agents invoking verified tools.
AI Model Risk Management Framework. Every Gate decision, tool invocation, and agent identity check is recorded in an Ed25519 hash-chained audit journal — tamper-evident provenance for inventory, accountability, and incident response.
Getting Real about AI Governance. Governance shifts from after-the-fact review to pre-execution enforcement. Cedar policies are version-controlled, statically analyzable, and produce auditable decision traces — the artifacts compliance and risk teams actually need.
Using AI for Offensive Security. The same runtime governs red-team agents: scoped capabilities, time-bound approvals, full attribution.
Open source. Standards-track via OATS. The policy engine, not the agent, holds the keys.
ThirdKey AI
Open-source trust stack for AI agents: typestate-enforced ORGA loop, Cedar policy gate outside LLM influence, signed tool schemas, cryptographic identity.
Learn moreTerra Security
Terra Platform is the agentic Offensive Security platform that replaces point-in-time pentests with fully integrated continuous attack-surface validation powered by agentic AI and paired with a Human-in-the-Loop. Every meaningful change to code, configuration, or reachable surface triggers evaluation. Terra maintains full context and proves what is actually exploitable as the environment evolves.
Terra Security
Terra Platform is the agentic, AI-powered continuous pentesting infrastructure that replaces point-in-time engagements with real-time, validated security testing aligned with every change to code and attack surfaces.
Learn moreKeel
AI agents can act across enterprise systems faster and more autonomously than traditional IAM and governance controls were designed for. Keel provides an independent authorization and evidence layer for consequential agent actions.
Before an agent acts, Keel evaluates the proposed action against policy and can allow, deny, require approval or challenge, or apply constraints such as budgets and scoped authority. Allowed actions can receive signed, scoped, time-bound AI Permits bound to the proposed action. Keel also supports delegated authority and revocation so permissions can be constrained as agents and sub-agents act across workflows.
Afterward, Keel produces tamper-evident cryptographic evidence of the authorized and enforced path that can be independently verified, including offline. This creates a durable link between agent identity, delegated authority, the requested action, the authorization decision, and execution evidence.
This directly addresses CSA’s call for fine-grained, dynamic authorization, traceable delegation, rapid revocation, and stronger accountability for autonomous agents. CSA’s Agentic AI IAM guidance identifies the need for task-specific permissions, real-time policy enforcement, clear delegation chains, and cryptographically verifiable audit records. Its Data Security guidance likewise emphasizes least privilege, policy-backed approval for high-risk actions, immutable audit trails, and fail-safe controls.
Keel complements existing IAM rather than replacing it: identity establishes who the agent is; Keel governs what that agent is authorized to do now and provides independently verifiable evidence of the authorization and enforced path afterward.
Keel
Keel controls consequential agent actions with signed AI Permits before they happen and produces independently verifiable evidence afterward.
Learn moreSurf AI
Surf AI is an agentic operations platform that closes the gap between knowing what's wrong and actually fixing it, continuously, safely, and at scale.
Modern enterprises face a structural problem: identity sprawl, overexposed data, and ungoverned AI agents create a growing attack surface that no single tool can see or act on. Surf addresses this directly.
On Identity and Access: Surf connects context across identity providers, HR systems, cloud platforms, and SaaS tools to surface risks that only become visible in combination. Dormant accounts, orphaned credentials, contractor access that outlasted employment, non-human identities with unchecked permissions. It traces every issue to a current owner, models downstream impact before any change is made, and uses specialized AI agents to drive resolution with human approval at every step.
On Data Security within AI Environments: As AI agents proliferate inside enterprises, they create new identities, new access paths, and new data exposures. Surf's context graph continuously maps these relationships so teams can enforce controls proactively rather than reactively.
On AI Security and Governance: Every action Surf takes is logged in a compliance-ready audit trail. Teams define policies, set guardrails, and retain approval authority over every decision. Surf operationalizes governance rather than treating it as a reporting exercise, ensuring that continuous security operations stay auditable, predictable, and safe.
The result is security hygiene that runs continuously, not in sprints, reducing exposure across identity, cloud, and data while keeping human teams in control.
Surf AI
Surf AI operationalizes your entire security program. AI agents find and close exposure gaps across identity, cloud, and data with your team in control.
Learn moreNoesiumIQ, Inc.
NoesiumIQ’s Cognitive Model Explainability Framework (CMEF) is a vendor-neutral AI governance and explainability SaaS platform. CMEF ingests AI and agent telemetry and transforms it into structured Evidence Packages for audit, assurance, compliance review, and responsible oversight.
In agentic environments, identity alone is not enough to establish accountability. Reviewers must be able to trace an action to the responsible agent, upstream user or agent delegation, model and version, relevant context, applicable policy or control state, decision output, and resulting action. CMEF is designed to assemble that chain into a reviewable Evidence Package without requiring organizations to replace their existing IAM, MLOps, security, or governance platforms.
Evidence Packages support five connected governance needs:
• Agent accountability: records the acting agent, delegation context, and responsible human or system boundary where that information is available through telemetry.
• Decision and action traceability: links decision telemetry to downstream actions, outcomes, and review events.
• Lifecycle governance: maintains evidence across model development, deployment, version changes, monitoring, incident review, and audit.
• Telemetry integrity: preserves provenance, timestamps, model and policy versions, control context, and evidence lineage so reviewers can identify both supported conclusions and evidence gaps.
• Evidence-based assurance: converts raw telemetry into structured, human-readable artifacts that security, compliance, risk, legal, and audit teams can review and challenge.
This aligns with CSA guidance on traceable agent identity and delegated actions, AI data protection and lifecycle governance, and governance maturity as a foundation for responsible AI adoption. CMEF is being developed toward an MVP and limited enterprise pilots, with roadmap support for NIST AI RMF, ISO/IEC 42001, EU AI Act-related governance needs, bias and drift documentation, and broader agentic AI oversight.
NoesiumIQ, Inc.
CMEF turns AI and agent telemetry into Evidence Packages for accountable decisions, traceable actions, lifecycle governance, and audit assurance.
Learn moreTenet Security
Tenet is the Agent Behavior Security Platform. It protects AI agents that take consequential actions and enforces the moment the agent acts.
The three CSA papers point to one gap. Agentic AI IAM shows OAuth, OIDC and SAML were built for predictable users and static apps, so they define permissions but cannot govern autonomous, machine-speed intent. The State of AI Security and Governance shows agents scaling faster than the controls around them, with 52% of leaders naming sensitive data exposure as their top risk. Data Security within AI Environments shows agents crossing data boundaries that perimeter controls never see.
Tenet closes that gap with a per-agent behavior profile: what the agent is allowed and disallowed to do, and what it could do with the keys it holds. We build it from inside the agent's own process, correlating OS activity, network and API calls, and LLM reasoning into one live timeline, with no heavy gateways, proxy or SDK. When the agent steps outside its profile, Tenet blocks the action in real time and can shut the agent down with a kill switch.
A permission can be correct and the action still wrong. At DEFCON last month we showed Ghostjacking: one planted alert walked a company's own agents into full cloud takeover, every step authorized, no control firing. Our Agentjacking research hit AI coding agents at 100+ companies the same way, both covered by Dark Reading, The Hacker News and Infosecurity Magazine. Only a behavior profile catches this.
This maps to the Agentic IAM framework: dynamic policy-based access control, continuous behavioral monitoring, non-repudiable logging, and rapid containment for agents. It is additive to identity and posture, not a replacement.
Tenet protects 24M+ end-users, has 5M+ agentic sessions analyzed, and has enterprise customers in production such as: LA Superior Court, GoPuff, Shift4 and PatientPoint.
Tenet Security
Tenet Security is the Agent Behavior Security Platform, we predict agentic misbehaviors and attacks before they happen.
Learn moreAstrix Security
AI agents now outnumber employees 100:1 at most enterprises. They connect to your most critical systems, operate autonomously, and run continuously, yet most fall outside existing IAM, audit cycles, and security review. Many have no human owner. Most have never been inventoried.
The core problem is access. Every AI agent operates through a non-human identity (NHI) that defines precisely what it can reach and what it can change. When those identities carry excessive permissions, remain unmonitored, or get compromised, the blast radius is significant and often invisible until after the fact.
Astrix is the platform built to close that gap, through a single framework: Discover, Secure, Deploy.
Astrix enables enterprises to:
-Discover every AI agent and NHI across cloud, SaaS, and on-premises environments, including shadow agents deployed without security review, using four methods: AI platform integrations, NHI fingerprinting, EDR sensor telemetry, and bring-your-own-sensor.
-Secure excessive privileges, misconfigured access, and policy violations, then remediates them. When agent behavior deviates from expected patterns, Astrix detects it in real time and surfaces it for response before it causes impact.
-And deploy agentic AI safely with secure-by-design guardrails. Agent policies enforce least-privilege access at the point of creation, with short-lived credentials and precisely scoped permissions. Security teams define what each agent is permitted to access and do before it runs, not after.
Most AI security approaches focus on model behavior, prompts, and outputs. Astrix operates at the identity layer, where agent risk is actually defined. We detect threats continuously, not just at provisioning. Most organizations are wrong about how many agents they have, typically by an order of magnitude. Finding them all, governing their access, and detecting when something goes wrong is the problem we were built to solve.
Astrix Security
Astrix discovers every AI Agent across the enterprise, controls what each agent can access and do, and detects threat before they make an impact
Learn moreaizome
aizome addresses the core challenge the CSA Agentic AI IAM framework identifies: existing identity protocols - OAuth 2.1, SAML, OIDC - were built for human users and static machine identities. They cannot govern enterprise AI agents that inherit user permissions, act autonomously across multiple systems, and can drift from their original purpose without human authorization.
aizome operates as a secured exchange between every AI agent deployed by the enterprise workforce and every enterprise system those agents access. Every agent is assigned a governed hybrid identity at onboarding - capturing its declared purpose, authorized systems, permitted toolset, and accountable human owner - directly addressing the CSA framework's call for rich, verifiable Agent IDs that go beyond static API keys or inherited credentials.
aizome implements the CSA framework's dynamic access control model: every agent action is evaluated against its declared purpose before execution, with four-layer policy enforcement covering session context, intent boundary, data protection, and behavioral drift detection. Permissions are scoped to the minimum required for the specific task - not the full permissions of the user who built the agent - directly addressing the CSA framework's data minimization and least-privilege requirements.
For multi-agent workflows, aizome preserves the original initiating user's identity and authorization context through every agent-to-agent handoff, preventing the accountability gaps and privilege escalation patterns the CSA framework identifies as a core risk of multi-agent systems. Unapproved cross-agent delegation is blocked at the boundary.
aizome's continuous behavioral monitoring establishes a baseline for every user-agent pair from onboarding and detects intent drift in real time across six dimensions - data volume, system reach, tool chain depth, outbound call rate, sensitivity tier, and instruction patterns - implementing the behavioral analytics and dynamic trust scoring the CSA framework prescribes. When drift is detected, the agent is contained before the deviation produces an incident.
Every agent interaction is logged in a tamper-evident audit trail from identity to tool call to action to outcome, structured for SIEM ingestion - satisfying the CSA framework's non-repudiation, accountability, and continuous monitoring requirements. Human-in-the-loop controls enforce explicit approval for consequential actions at the infrastructure layer, not the prompt layer - implementing the structural human oversight the framework requires.
aizome deploys into the customer's AWS VPC with no agent modification required, integrating with existing IdP infrastructure via SAML/SSO and connecting to enterprise systems including Salesforce, SAP, ServiceNow, and on-premises applications. Discovery results are visible within hours of deployment.
aizome
aizome is the AI Identity Gateway - governing identity, intent, and cost for every AI agent your workforce deploys across enterprise systems.
Learn moreAkto
Akto secures the full agentic AI estate: every homegrown AI application, agent, LLM, and MCP server your organization builds, and every AI tool, copilot, and shadow AI app your employees use.
- Discovery & Observability: 50+ connectors across agent builder platforms (Copilot Studio, Bedrock, Vertex AI, Snowflake, LangChain, Anthropic, and more ), AI gateways (LiteLLM, Portkey), and cloud infrastructure delivering live inventory of every AI agent, MCP server, LLM, and RAG database with Context Graph visualization. On the employee side, browser extensions, Endpoint Shield, and IDE hooks surface shadow AI usage.
- Identity & Governance: Inventories every credential an agent touches from API keys, OAuth tokens, to service account keys tied to the agent, and the human accountable for it, and its access scope. Policies enforce least-privilege and block cross-service reuse. Violations are logged with severity, blast radius, and remediation steps.
- AI Red Teaming: Automated adversarial testing with 4,500+ probes covering prompt injection, jailbreaks, indirect prompt injection, and multi-turn attacks. Covers OWASP Top 10 for LLMs, Agentic Applications, and MCP Servers, including a dedicated MCP scanning module.
- Runtime Protection & Guardrails: AI-powered and policy-based guardrails for real-time blocking; anomaly detection, agent abuse, and rogue agent detection. Enforces policies controlling agent actions, restricting data exfiltration, tool calls, harmful topics, hallucinations, and prompt injection. Inline blocking via Akto AI Agent Proxy, customer proxy integration, or alert-only mode.
- AI Security Posture Management: Continuously assesses risk across every AI app, scanning configurations, identifying policy gaps, surfacing misconfigurations, and tracking exposure across third-party dependencies.
Akto supports both SaaS and on-prem. Customers can use frontier models, Akto model or bring their own model for AI capabilities.
Akto
Akto is the agentic AI security platform securing AI agents, MCP servers, LLMs, and employee AI usage through discovery, red teaming, and runtime enforcement.
Learn moreSecurity Consultant OÜ
Security Consultants helps organizations secure, govern, and operationalize agentic AI. We translate AI governance, security, privacy, and compliance requirements into practical controls and working processes across the full lifecycle, including risk assessment, data governance, human oversight, vendor risk, monitoring, and audit readiness. Our work helps enterprises adopt autonomous AI systems with greater trust, resilience, and regulatory confidence.
Security Consultant OÜ
Security Consultants helps organizations secure AI and cloud operations through vCISO, audits, compliance, and certification support.
Learn moreKontext
Kontext provides runtime control and authorization for AI agents by checking every supported Claude Code, Codex, and Cowork tool call against policy before it runs. As enterprises move from assistants that generate text to agents that execute commands, edit files, and call APIs, authentication alone is not enough: individually permitted actions can be chained into outcomes nobody intended.
Kontext adds an independent enforcement layer between agent intent and execution. It evaluates each proposed action using deterministic policy and execution context, allows, denies, or escalates the action before the side effect, and records a normalized decision fact for auditability. A separate local cyber-risk signal can inform uncertain cases, but the enforcement guarantee remains policy-based rather than relying on the model to police itself.
This maps most directly to CSA's Agentic AI Identity and Access Management guidance: authority should be task-scoped, least-privilege, revocable, and enforced at action time. It also supports CSA's AI governance and data-security objectives by giving security teams transparent controls and evidence for how agents interact with sensitive systems. Kontext is designed to remain vendor-neutral at the policy layer and currently supports Claude Code, Codex, and Cowork surfaces.
For defense in depth, Kontext's open-source Sandy project adds fail-closed native process containment on macOS and Linux. Together, runtime authorization and process containment help preserve intent-to-execution integrity without treating prompts or post-hoc monitoring as a security boundary.
Kontext
Runtime authorization for AI agents that checks supported tool calls against policy before execution and records evidence for every decision.
Learn moreAbove Security
Above is a proactive insider risk platform built on a fleet of AI investigative agents. It directly embodies the agentic-AI patterns CSA describes: autonomous agents that reason over context, invoke tools through a governed gateway, and act within scoped, least-privilege boundaries rather than static rules.
Where legacy detection relies on brittle policies and after-the-fact alerts, Above's agents continuously investigate identity behavior — correlating signals across SaaS, internal, and custom applications with enrichment from IdP, HRIS, EDR, and email-security sources. Each investigation is an autonomous, multi-step reasoning task: the agent gathers evidence, forms and tests hypotheses, and produces an explainable verdict a human analyst can audit.
Above's architecture reflects CSA's guidance on secure, trustworthy agentic design:
Governed tool use: agents call live "runtime" tools only through a controlled gateway (AgentCore), so every action is mediated, scoped per-tenant, and observable — never unbounded code execution.
Least privilege and isolation: multi-tenant scoping and identity-visibility boundaries constrain what any agent can see or do, aligning with agentic IAM and non-human-identity principles.
Human-in-the-loop and explainability: investigations yield transparent, cited reasoning, keeping a person in control of consequential decisions.
Defense for the agentic era: as organizations deploy their own AI agents and non-human identities, Above extends behavioral investigation to those actors — surfacing misuse, compromised credentials, and anomalous agent activity that traditional tools miss.
Above both applies agentic AI responsibly and helps enterprises secure the expanding identity surface that agentic AI creates — turning insider risk from a policy problem into a continuously investigated one.
Above Security
Above is a proactive insider risk platform using a fleet of AI investigative agents to continuously investigate identity behavior and surface real threats.
Learn moreKeter Labs
Traditional IAM (OAuth 2.1, SAML) is fundamentally broken for autonomous agentic workflows: as highlighted in the CSA's 'Agentic AI Identity & Access Management' framework, static bearer tokens cannot evaluate multi-step action trajectories, prevent confused deputy attacks, or operate at machine speed.
Reality Kernel by Keter Labs is the industry's first operational implementation of the CSA's Layer 4 Adapter Enforcement Middleware (AEM) and pre-dispatch Policy Enforcement Point. Operating directly in the execution runtime (Python SDK v0.7.1 on PyPI), Reality Kernel solves the core architectural gaps identified in the CSA publication:
1. Deterministic Pre-Dispatch Containment: Rather than relying on slow, probabilistic LLM-as-a-judge guardrails (which add 800ms+ network latency and remain vulnerable to prompt injection), Reality Kernel computes intent divergence and validates tool execution boundaries with sub-millisecond local evaluation, intercepting anomalous or hijacked calls before network sockets or system commands are dispatched.
2. Cryptographic Non-Repudiation & Auditability: Aligning directly with the CSA framework's specification for EdDSA/Ed25519 verifiable execution tokens, Reality Kernel cryptographically seals every agent dispatch with canonical context hashing. This creates an immutable, tamper-proof audit trail mapped directly to EU AI Act Article 14 and ISO 42001 requirements.
3. Zero-Breakage Enterprise Adoption: To solve the deployment friction noted in Section 6, Reality Kernel features an enterprise Shadow Mode. Security teams can monitor, baseline, and cryptographically audit autonomous agent fleets in production with zero impact on latency or workflow reliability before enabling active enforcement.
Reality Kernel transforms agent authority from an unverified string into a mathematically verifiable, sub-millisecond runtime trust barrier.
Keter Labs
Sub-millisecond cryptographic pre-dispatch barrier for AI agents. Intercepts untrusted tool calls and enforces deterministic runtime trust before execution.
Learn moreLumia Security
The CSA report "The State of AI Security and Governance" identifies governance maturity as the strongest predictor of successful AI adoption, while highlighting data exposure, compliance requirements, shadow AI, and agentic AI as the primary challenges organizations face. Lumia directly addresses these gaps by providing a real-time governance and control layer for both employees and autonomous agents. Rather than relying on policies alone, Lumia continuously assesses AI interactions based on content, context, and intent, giving organizations visibility into AI usage and the ability to enforce security, business, and regulatory policies before sensitive information is exposed or actions are executed.
The report further concludes that traditional security controls are insufficient for managing the behavioral and non-deterministic nature of AI systems. Lumia was purpose-built for this challenge. By understanding the meaning of prompts, responses, files, and agent actions, Lumia enables adaptive, real-time governance across AI applications, models, and agents. This helps organizations reduce AI-related risk, maintain compliance, prevent oversharing, and securely scale AI adoption while keeping employees and autonomous agents aligned with organizational policy.
Lumia Security
Lumia governs how employees and AI agents use AI, ensuring every interaction aligns with business policy, security requirements, and compliance obligations.
Learn moreKonaSense
KonaSense is a people first AI Security platform built to protect how employees use AI at work. We deliver real time visibility and control across the full AI workflow: employee interactions in the browser, copilots and chat tools, and agentic systems running through KonaProxy. That includes prompts, uploads, copy and paste, tool calls, and model routing. We enforce policy in the moment with actions like allow, block, redact, and coach so risky behavior is stopped before data leaves your environment.
KonaSense unifies AI Governance, AI Security, and AI Observability into one control layer. The platform continuously analyzes AI activity, detects sensitive data exposure, shadow AI, and unauthorized model usage, and produces audit ready evidence.
KonaSense
KonaSense secures and governs employee and agentic AI use with real time visibility, policy enforcement, redaction, and audit ready controls.
Learn moreExeclave
Execlave is a runtime enforcement and evidence layer for AI agents. Before an agent calls a tool, sends data or moves money, it asks Execlave with one SDK call. Execlave checks the action against the organization's policy and returns PASS, HOLD for human approval, or DENY before the action runs. Measured in production, the server-side decision takes 19.9 ms at the median.
The State of AI Security and Governance: Execlave turns governance policy into enforced controls. 19 built-in policy types cover tool access, spend limits, PII, model restrictions and prompt-injection guards. Each agent gets one of four autonomy levels (observe, advise, act-with-approval, autonomous) with a matching policy bundle. Blocks, holds and approvals are written to a hash-chained, append-only audit trail sealed by signed checkpoints, with evidence mapped to SOC 2, ISO 27001, EU AI Act, GDPR, HIPAA, PCI DSS, NIST AI RMF and OWASP ASI.
Agentic AI Identity and Access Management: Execlave enforces least privilege per agent. An agent registry tracks every agent and its versions, shadow-agent detection finds unregistered agents, and permission-drift detection flags privilege escalation and unused, over-privileged permissions. A kill switch pauses a single agent so its next action is blocked.
Data Security within AI Environments: PII is detected and redacted before storage by default, and data-access lineage records which data classes (PII, PHI, PCI, confidential) each agent touched.
Agent supply chain: MCP tool descriptors are pinned per agent and checked on every call, so a changed or poisoned tool is blocked before the agent uses it.
Execlave is framework-neutral (LangChain, CrewAI, OpenAI Agents SDK, AutoGen, LlamaIndex, MCP; Python and TypeScript SDKs) and runs as managed cloud or self-hosted, including air-gapped.
Execlave
Execlave checks every AI agent action against policy before it runs: allow, hold for approval or block, with a tamper-evident audit trail. Cloud or on-prem.
Learn moreLangGuard Inc
What sets LangGuard apart from other runtime tools is that the policies carry business context. Most runtime security products enforce generic rules, such as blocked tools, PII filters, and rate limits. LangGuard connects to the enterprise AI governance layer, where risk appetite, approved use cases, and accountable owners are defined. It turns those definitions into enforceable runtime policy. Our policy authoring agent converts governance intent into Rego without code, and GRAIL adds context from systems of record. As a result, Arbiter can tell the difference between "this agent can call the payments API" and "this agent may not approve a payment it also initiated, above the threshold the Controller set." The governance team writes the policy once, and LangGuard enforces it on every action.
LangGuard Inc
What sets LangGuard apart from other runtime tools is that the policies carry business context. Most runtime security products enforce generic rules, such as blocked tools, PII filters, and rate limits. LangGuard connects to the enterprise AI governance layer, where risk appetite, approved use cases, and accountable owners are defined. It turns those definitions into enforceable runtime policy. Our policy authoring agent converts governance intent into Rego without code, and GRAIL adds context from systems of record. As a result, Arbiter can tell the difference between "this agent can call the payments API" and "this agent may not approve a payment it also initiated, above the threshold the Controller set." The governance team writes the policy once, and LangGuard enforces it on every action.
Learn moreAegisGate Security, LLC
AegisGate is an open-source AI security gateway that implements zero-trust identity, policy-based access control, and continuous monitoring for agentic AI systems. The platform sits between AI agents and the services they call, inspecting traffic across five protocols (MCP, A2A, ACP, ANP, HTTP) using a three-layer detection pipeline: 216 regex patterns, MITRE ATLAS threat mapping, and a CharCNN-BiLSTM ML classifier.
Identity: Four agent roles (restricted, standard, privileged, admin) with hierarchical permissions, a tool authorization risk matrix requiring human approval for high-risk operations, SAML/OIDC SSO with JWKS rotation, and multi-tenant isolation via PostgreSQL Row-Level Security.
Governance: 31 compliance frameworks implemented as automated CheckFuncs, including NIST AI RMF, ISO 42001, and EU AI Act — three of the five frameworks the CSA AICM itself cross-references. OPA/Rego policy engine, RFC 5424 audit logging with cross-framework reference mapping (ATLAS, NIST, OWASP LLM, CWE, CVE), and cryptographically signed evidence packages.
Observability: Anomaly detection, real-time monitoring, incident response with MITRE ATLAS enrichment, SOAR integration, STIX/TAXII IOC ingestion, and a dedicated CVE-for-AI feed with tamper-evident attestation.
Supply Chain Integrity: AIBOM (CycloneDX 1.6) — a fully implemented generator, signer, and verifier for AI Bill of Materials covering five protocol pillars, models, prompts, and RAG corpora. ECDSA P-256 attestation envelopes make every claim tamper-evident and third-party-verifiable.
Vendor-Neutral Guardrails: Protocol-specific guards for MCP, A2A, ACP, and ANP enforce rate limits, session limits, tool call limits, and execution timeouts. Response guard inspects AI output for PII, secrets, toxicity, hallucination, and token abuse before it reaches users or context stores.
AegisGate Security, LLC
Open-source zero-trust AI security gateway: agent RBAC, multi-protocol guardrails, AIBOM, and 31 compliance frameworks for agentic AI traffic inspection.
Learn moreThe Cyber BEAR Group Inc. (CBG) GRIZZLY
Grizzly is the definitive Asset Intelligence Platform built to reclaim control of agentic workflows and "dark data". Grizzly is the single source of truth for all digital Assets. A cohesive, automated, holistic platform making every digital Asset instantly discoverable, accounted for, secured, policy and regulation enforced, throughout the Asset’s lifecycle. Through our AI native, digital Asset management platform, we deliver:
Identity & Observability: We resolve the modern "Identity Crisis" by cryptographically securing and managing human and Non-Human Identities (NHIs) and tracking AI agent artifacts with enforcement down to the field level. We provide total observability by mapping data lineage, bridging the 40%-60% visibility gap where shadow data typically hides.
Supply Chain Integrity: Grizzly prevents agentic attacks and vulnerabilities by binding every piece of code and build-artifact to a verifiable identity and policy. We treat SBOMs and container images as signed assets, ensuring that any tampering within the CI/CD pipeline is immediately detectable.
Vendor-Neutral Guardrails: Our agnostic, agentless platform doesn't replace your stack; it completes it. We provide universal guardrails that work across any AI model or cloud provider, preventing vendor lock-in while enforcing consistent policy across your entire digital landscape.
Ironclad Governance & Effortless Compliance: We shift the burden from reactive audits to being Continuously Audit Ready. Our platform maintains an unbreakable, tamper-evident chain of custody and forensic provenance, enabling "One Touch Reporting" for boards and regulators.
Machine-Speed Response: When threats strike, Grizzly enables asset-centric containment in minutes, not days. Security teams can instantly assess breach scope, revoke rogue agent keys, and freeze compromised assets at the speed of the attack.
Grizzly’s full spectrum Asset integrity provides business and operational certainty that is required to innovate at the speed of AI.
The Cyber BEAR Group Inc. (CBG) GRIZZLY
CBG's GRIZZLY is an AI Native Asset Management platform, fortifying the future of digital assets. We integrate your entire business environment, automate compliance and operations, providing the freedom to innovate without the fear of what's lurking in your blind spots. PPIO
Learn moreMetricsLM
AI agents change after approval. Their tools, permissions, data access, integrations and autonomy can expand while the original assurance decision remains unchanged.
MetricsLM continuously answers one question: Is the AI system running today still the system the enterprise approved?
MetricsLM establishes an approved operating boundary for an AI system, covering its purpose, models, data, tools, permissions, integrations and autonomy, and connects that boundary to technical evidence from sources such as code repositories and identity systems.
Once approved, MetricsLM baselines the system and monitors for material change. If an agent gains a new permission, tool or data source, receives write authority, increases its autonomy, or removes a human approval step, MetricsLM identifies the change against the approved boundary and triggers reassessment with an auditable evidence trail.
This creates an assurance layer alongside runtime security and IAM. Runtime controls ask whether an individual action should be allowed. MetricsLM asks whether the system running today is still operating within the boundary the enterprise originally reviewed and approved.
As autonomous agents become increasingly dynamic, MetricsLM turns AI assurance from a point-in-time review into a continuous process: define the boundary, baseline it, detect material change, reassess.
MetricsLM
MetricsLM continuously verifies AI systems against their approved boundary, detects material change and triggers reassessment. Audit-ready. Always-ready.
Learn moreHAVENLON
HAVENLON builds an independent Execution Control Layer for autonomous AI: a security boundary between an agent's authorized intent and the real-world action it can cause. Our premise is that valid identity, permissions, and approval do not necessarily guarantee correct execution—an "Execution Gap" that grows as agents gain access to payments, assets, data, and critical systems.
In line with CSA's Agentic AI Identity and Access Management guidance, HAVENLON binds signed intent and authorization to the specific action, target, parameters, current state, and required evidence at the execution boundary. It complements existing IAM rather than replacing it. Missing, expired, conflicting, or unverifiable conditions lead to denial, not implicit permission.
To address the governance concerns highlighted in CSA's State of AI Security and Governance, participants can independently define constraints; the effective boundary is their intersection, not a majority vote or administrator override. Bletchley SaaS coordinates policies but does not hold final execution authority. Cryptographically linked evidence records support review of execution decisions and outcomes.
Relevant to CSA's Data Security within AI Environments, these controls can constrain high-risk agent-initiated data transfers and system actions when integrated into the execution path; HAVENLON is not presented as a general-purpose DLP tool.
One architecture is designed for two assurance levels: Enigma dedicated, customer-controlled security hardware (working engineering prototypes) and a software-only execution boundary (under development). The approach is designed to be vendor-neutral and non-custodial. It is grounded in nearly two years of engineering research, three arXiv preprints, and the three-volume Execution Engineering Trilogy.
Control before execution. Proof after execution.
HAVENLON
HAVENLON secures autonomous AI execution through software and hardware boundaries that enforce policies before actions and produce verifiable evidence.
Learn moreAdversa AI
The Cloud Security Alliance warns that AI threats evolve faster than traditional security controls. Static guardrails and one-off pentests age quickly as models drift and agentic workflows change. In AI, “set and forget” security becomes a liability.
Adversa AI is an autonomous, continuous red-teaming platform for your AI stack. We don’t just help you deploy AI safely once. We continuously attack your models, guardrails, and agent workflows to validate that your defenses hold up against real-world threats.
How we deliver on CSA’s mandate for dynamic model assurance:
1️⃣ Continuous guardrail validation: The CSA stresses the need to test data detectors and tokenizers using dynamic "red-team prompts". Adversa AI replaces point-in-time testing with a continuous attack engine that invents novel, multi-step jailbreaks and tool-abuse attacks, validating that your guardrails catch zero-day threats.
2️⃣ Business-context exploitation: We don't run generic payloads. We ingest your specific business context (fintech privacy rules or internal API structures) and craft domain-specific exploit chains to uncover logic-based vulnerabilities that matter in your environment.
3️⃣ Instant framework alignment: Every finding is mapped to MITRE ATLAS, OWASP LLM Top 10, NIST guidance, and EU AI Act categories. Security and compliance teams get audit-ready reports instead of ad hoc testing notes.
4️⃣ Developer-friendly remediation: We don't just break your AI. We help you fix it. Adversa integrates with CI/CD pipelines and task trackers, providing “autopatch” suggestions, hardening playbooks, and concrete guardrail improvements your engineers can apply immediately.
With flexible deployment options (including fully air-gapped, on-prem engines where your data never leaves your environment), Adversa AI turns the expensive, slow process of manual red teaming into a continuous, automated security control.
Innovate with confidence. Break your AI before attackers do.
Adversa AI
Adversa AI is an autonomous, continuous red-teaming platform that secures your AI stack by validating guardrails and testing against your unique business risks.
Learn moreSurePath AI
In today’s AI-everywhere world, organizations face a critical paradox: allow transformative technology with significant risks or fall behind. AI’s adoption has created a new layer of enterprise risk that traditional security tools were not designed to handle: employees use public AI tools, approved or not, in browsers; developers rely on copilots and local agents; and AI is embedded into internal applications. Increasingly, autonomous agents are acting on behalf of users, accessing systems, chaining tools, and making decisions in real time.
SurePath AI addresses the challenges by operating as a network-level control plane for AI. Through real-time interception, model routing, and policy enforcement, SurePath AI provides full visibility and control without slowing teams down. It secures AI access, restricts interactions to approved systems, detects malicious or risky behavior, and protects sensitive data before it reaches a model. SurePath AI ensures AI accelerates business outcomes, not risk.
Key Capabilities
Visibility & Auditability: Capture and monitor AI usage across the organization, including prompts, responses, and tool interactions. Comprehensive audit trails provide a clear record of who used AI, how it was used, and what actions were taken.
Insights & Analytics: Turn AI activity into actionable insights. Identify usage patterns, understand user intent, and generate reports for security, governance, and compliance teams.
Data Protection: Detect and redact sensitive information in real time, preventing data leakage while preserving user productivity.
Access Control: Integrate with directory services to enforce role- and group-based access to models, tools, and enterprise data.
Policy Enforcement: Apply granular, policy-driven guardrails to allow, block, redact, or monitor AI interactions ensuring safe, compliant AI adoption at scale.
SurePath AI
SurePath AI governs AI at the network layer across public and private models, AI coding assistants, homegrown applications, and autonomous agents.
Learn moreToken Security
Token Security believes that AI agents are entities distinct from traditional human or machine identities, “capable of planning, reasoning, and executing complex, multi-step tasks with minimal human supervision”.
The central finding of “Agentic AI Identity and Access Management: A New Approach" is that the protocols we have are not capable of governing them. At the same time, legacy IAM tools can’t keep up with the autonomy, ephemerality, and delegation patterns of AI agents in complex multi-agent systems. We agree. Identity is the only control plane capable of securing and governing agentic AI. Prompt filtering comes too late to be relied upon for security because it doesn’t control an AI agent's access. The confused deputy problem described in the paper is what we see in nearly every enterprise we assess.
Token Security believes that intent-based security is required for AI agents. An agent's declared purpose becomes the boundary against which its access and actions are judged. Our platform:
* Discovers every AI agent and the non-human identities it uses across cloud, SaaS, endpoints, and MCP servers.
* Understands each agent's access, behavior, owner, and intent.
* Enforces least privilege scoped to that intent, with lifecycle control from creation through decommissioning.
We do this by creating an AI agent inventory, analyzing contextual data, providing full AI agent lifecycle management, AI Agent Security Posture Management, AI Threat Detection & Response, Access Controls, and automated policy enforcement at scale.
The lifecycle gap the paper flags is measurable. In CSA's own survey with Token (https://cloudsecurityalliance.org/press-releases/2026/04/21/new-cloud-security-alliance-survey-reveals-82-of-enterprises-have-unknown-ai-agents-in-their-environments), 82% of organizations had found a previously unknown AI agent, and only 21% had a formal decommissioning process. Dormant identities holding live credentials are the liability Section 3 anticipates.
Token Security
Token Security accelerates the secure adoption of agentic AI by discovering AI agents, understanding their access and intent, and enforcing security controls.
Learn moreMarquor
As AI agents become increasingly autonomous, enterprises need more than agent identity they need to know whether an agent has the authority to perform a specific action at runtime.
Marquor is a payload-free runtime authorization and governance layer for AI agents. It provides an independent control point that evaluates agent actions against enterprise-defined authority and governance policies before those actions are allowed to proceed.
Marquor addresses a critical challenge in agentic AI: an authenticated agent may still attempt an action outside its intended authority, delegated permissions, or governance boundary. Marquor continuously evaluates these interactions at runtime, allowing authorized actions while denying or escalating actions that violate established boundaries.
Unlike approaches that require inspection of prompts, business data, or application payloads, Marquor is designed to operate payload-free. This enables organizations—particularly those operating in regulated and privacy-sensitive environments to enforce agent authorization without exposing sensitive customer or enterprise data to the authorization layer.
Marquor supports CSA's vision for Agentic AI Identity and Access Management by extending access control beyond establishing agent identity toward runtime authority enforcement: determining whether an identified agent is authorized to take a particular action in a particular context.
The result is a vendor-neutral governance and authorization layer designed to help enterprises deploy autonomous AI agents with stronger control, privacy, accountability, and confidence.
Marquor
Payload-free runtime authorization and governance for AI agents, enforcing whether an agent has authority to perform an action before execution.
Learn moreIlluminAIT
IlluminAIT — Comprehensive AI Visibility
Leadership is pushing to deploy AI now. Employees are already using it — in developer-built tools, vibe-coded MCP servers, and unsanctioned SaaS integrations that never touched a security process. The honest question is: does anybody actually know where AI is running in their network — and critically, what data it's touching?
Most don't. Only 26% of enterprises have comprehensive AI governance policies. 68% cannot distinguish AI agent activity from human activity. 72% lack confidence in their ability to execute a security strategy for AI they've already deployed. Governance policies are necessary, but AI is now so simple to deploy that policy alone cannot keep pace. Across CSA's thought leadership on agentic AI security, one core prerequisite is apparent: AI security, governance, and agentic identity depend on knowing what AI systems exist. That foundational AI inventory is what most organizations are missing.
IlluminAIT delivers it.
Our automated, agentless platform discovers AI workloads - models, agents, MCP servers, and the data sources they access - including shadow deployments. Unlike tools that infer AI presence from logs or metadata, IlluminAIT performs primary discovery in the network, building a deterministic runtime connectivity graph that maps the complete agentic workflow: model → agent → tool → MCP → data source. The result is a continuously updated, authoritative AI asset inventory with contextualized network topology visualizations - answering the questions governance frameworks can't: What AI is running? What is it connected to? What data can it reach? This foundational visibility layer makes shadow AI detection, agentic identity governance, data flow documentation, and AICM compliance all actionable.
Responsible and effective AI adoption requires proactive verification in the form of comprehensive discovery, visibility, and inventory of AI workloads running across the enterprise.
You can’t secure what you can’t see. IlluminAIT: comprehensive AI visibility.
IlluminAIT
Comprehensive agentic visibility: automated, agentless scanning discovers AI workloads in Enterprise networks - models, agents, MCP servers, accessible data sources.
Learn moreNuDay AI
NuDay offers a true zero trust approach to agentic AI. We are a complete platform, including management and control plane, the agents themselves, and runtime execution layer. We dynamically encrypt in real-time the entire history of the running agents. We also have a unique vector obfuscated, data encrypted RAG solution to protect RAG data attached to agentic systems. All of the tools, guardrails, skills, and even agent configuration are digitally signed with a PKI check before every invocation, thereby preventing tampering or execution of unsigned code or commands. Our agent tracing and audit are also cryptographically signed making them tamper-evident and ideal for regulated environments.
NuDay AI
Zero trust enterprise agentic AI security, governance, and runtime platform. We encrypt agent history and RAG data in realtime and cryptographically enforce agent actions.
Learn morePillar Security, Inc.
AI agents now run everywhere: in the apps enterprises build, in the SaaS they buy, and on every developer's workstation. Pillar secures all three in one platform, where business context connects discovery, testing and protection so security intelligence compounds across the lifecycle.
The Discovery & Posture pillars inventory every agent, model, prompt, MCP server, skill and tool, including shadow AI, and flag wildcard permissions, unapproved PII and production access, and missing human-in-the-loop gates: the least-privilege failures CSA's Agentic AI IAM paper warns about.
The Agentic Endpoint pillar reaches where most shadow AI lives. Developers run coding agents that execute commands, chain tool calls, and reach production through MCP servers, unseen by security. Pillar discovers every coding agent, MCP server, skill, plugin, and configuration file across the fleet, analyzes credentials, access scopes, and auto-run settings, and monitors every prompt, tool call, and command for prompt injection, tool poisoning, and exfiltration.
RedGraph, Pillar's agentic red teaming engine, maps that environment as an attack graph and tests it continuously: attacker agents chain multi-turn attacks, indirect prompt injection and tool hijacking into proven exploits, each documented with a full transcript and reproduction steps. RedGraph also black-box tests AI features embedded in SaaS apps, probing shared-content injection, over-exposed data, and access-control gaps.
Runtime Guardrails, calibrated to each agent's business purpose rather than generic anomaly baselines, deliver precise controls. Taint analysis traces PII, credentials, and secrets from source to destination and blocks unauthorized egress in real time. Governance enforces model and MCP allowlists and maps evidence to the OWASP Top 10s for LLM and Agentic Applications, MITRE ATLAS, NIST AI RMF, and ISO 42001.
Everything deploys self-hosted in your own cloud, with SOC 2 Type II: the purpose-built practice CSA's State of AI Security and Governance report says securing AI now requires.
Pillar Security, Inc.
Pillar's AI security platform: agent discovery and posture, endpoint/coding agent protection, agentic red teaming, runtime guardrails and compliance governance.
Learn moreHOL
AI coding agents (Claude Code, Codex, Cursor, Gemini CLI) now act with the developer's own permissions: they run shell commands, read files that hold secrets, call MCP tools, install packages and load plugins. In line with the agentic IAM theme of giving autonomous agents scoped, reviewable permissions rather than unchecked inherited access, HOL Guard applies that idea at the workstation. Developers write a policy once; before each agent action Guard checks it locally and answers allow, block or ask, so sensitive operations need an explicit human yes and secrets or risky installs can be refused outright. Decisions are made on the developer's machine, independent of any one model or agent vendor, which suits vendor-neutral guardrails and governance. The core is an open-source Apache-2.0 CLI (pipx install hol-guard && hol-guard init); Guard Cloud is an optional hosted, managed paid tier for teams. Honest scope: Guard is not a gateway, not a network firewall and not a complete prompt-injection defence; it governs what a local coding agent is allowed to do.
HOL
Local-first rule check for AI coding agents: shell, secret, MCP, package and plugin actions are allowed, blocked or held for approval on the developer machine.
Learn moreStraiker
Straiker is The Agentic Security Company.
AI agents are shipping faster than security teams can inventory them, test them, or protect them at runtime. Straiker solves all three.
Discover AI maps every AI app, agent, MCP server, and model running across your organization and grades the risk. Ascend AI continuously red-teams them with real-world adversarial prompts across 30+ exploit categories. Defend AI enforces policy at runtime, catches novel attacks with ML, and alerts your SOC. Every attack Ascend AI finds becomes a Defend AI policy recommendation. The loop closes.
Straiker's adversarial dataset was built through pre-release engagements with leading frontier model labs, producing visibility into agentic failure modes before they enter public disclosure. Detection runs at p50 latency of 127ms, 98.1% accuracy, and 6 to 21x lower false-positive rates than prompted frontier LLM judges.
Customers include Comcast, Snowflake, and AMEX GBT which runs Defend AI at 10 million agentic requests per day. On-premises Helm deployment is available for regulated environments.
Straiker was founded by CEO Ankur Shah (Prisma Cloud, $500M ARR at Palo Alto Networks) and CTO Sreenath Kurapati (Cyberfend, acquired by Akamai). The STAR Research team includes practitioners from FireEye, Mandiant, Microsoft, Meta, xAI, and the U.S. Department of Defense.
Straiker
Agentic security from inventory to runtime. Discover AI finds every agent. Ascend AI red-teams them. Defend AI stops attacks.
Learn moreAstraSync AI Pty Ltd
CSA's Agentic AI IAM paper shows traditional IAM breaks when agents act autonomously, spawn sub-agents and act on delegated authority, and asks who is accountable when a sub-agent misuses data. AstraSync answers that in production. Every agent is bound to a verifiable five-party trust chain: developer, owner, agent, instructor and the counterparty relying on it. No unbroken human chain, no approval to operate.
We implement the paper's core components:
- Agent ID anchored to DIDs and verifiable credentials, with provenance linking every sub-agent to accountable humans
- Dynamic PBAC: PDLSS policies (Purpose, Duration, Limit, Scope, Self-instantiation) enforced at runtime with just-in-time narrowing
- Continuous monitoring: a 0-100 trust score flags behavioural drift
- Non-repudiation: tamper-evident audit trails with ledger-anchored proofs
Data Security within AI Environments calls for new containment as agents access data across boundaries. AstraSync operationalises the agentic clause of AICM DSP-17, monitoring, controlling and logging every agent and NHI access to sensitive data. Instant revocation provides the kill-switch the paper recommends for critical sectors, and surfacing unregistered agents supports proposed DSP-28 on shadow AI. Both papers note ledger trade-offs, so our design is hybrid: zero-gas anchoring of proofs only, personal data held off-chain to respect erasure rights, and developer verification through accredited identity providers.
The State of AI Security and Governance found only 26% of organisations have comprehensive AI governance, and those that do are twice as likely to adopt agentic AI. AstraSync turns governance from a policy document into controls that execute on every agent action.
Neutral by design, we layer over existing IdPs across MCP, A2A and agentic commerce, serving both the enterprise deploying agents and the counterparty receiving them.
AstraSync AI Pty Ltd
Neutral Know Your Agent infrastructure: binds every AI agent to verified humans, enforces what it may do at runtime, and proves what it did.
Learn morePointGuard AI
PointGuard AI directly bridges the critical gaps identified in the CSA State of AI Security and Governance 2025 and Data Security within AI Environments research papers. While CSA research reveals that 60% of enterprises are adopting agentic AI, yet 72% lack confidence in securing it, PointGuard AI provides an enterprise-ready platform designed specifically for the autonomous AI era. To solve the top industry challenge, sensitive data exposure (52%), PointGuard AI delivers AI-native Data Loss Prevention and real-time prompt protection to halt data leaks and indirect prompt injection attacks. Addressing the core governance maturity gap highlighted by CSA, our platform operationalizes the CSA AI Controls Matrix (AICM) through complete multi-vector AI discovery, shadow AI detection, AI-BOM inventories, and automated compliance reporting. Furthermore, PointGuard AI directly fulfills CSA's proposed control extensions (DSP-25 through DSP-28) by governing agent-to-tool ecosystems via our MCP Security Gateway and "Risk Rubric" Knowledge-base (collaborated with CSA). Finally, PointGuard AI establishes zero-trust NHI management, OAuth & OBO token exchanges, and active runtime containment through Agent Mission Control, complete with agent kill-switches, and human-in-the-loop oversight. PointGuard AI transforms non-deterministic AI risks into managed, compliant enterprise adoption.
PointGuard AI
PointGuard AI delivers total visibility, governance, and runtime protection for Agentic AI via security gateways, DLP, guardrails, and kill switches
Learn morePermiso Security
The AI agent skill supply chain broke in public in early 2026. Within six weeks of Anthropic publishing the Agent Skills open standard, security researchers found 341 confirmed malicious skills on ClawHub — a 12% infection rate. Not one was flagged by a static scanner before deployment.
The problem isn't that organizations aren't checking. It's that static analysis — whether automated rules or LLM code review — can only evaluate what a skill says it does. It cannot tell you what it does when a live agent runs it in a real environment. Malicious skills are designed to pass that test.
SandyClaw addresses this gap directly. Built by the Permiso P0 Labs team — the same researchers who found malicious skills on the marketplaces before most organizations were paying attention — SandyClaw is a dynamic detonation platform for AI agent skills. You submit a skill, a prompt, and a model. SandyClaw spins up a sandboxed agent environment, executes the skill, and records everything: every LLM action, every network call and domain resolution, every file write, SSL-intercepted and decrypted traffic, environment variable access attempts, and webhook traffic. Analysis runs against Sigma, Yara, Nova, and Snort — augmented with custom detection rules from P0 Labs. You get a verdict, a full findings list, and the complete behavioral record behind every determination.
This is the same runtime-versus-static distinction that has defined security analysis for two decades — the same philosophy that displaced signature scanning for malware. The agent skill ecosystem is currently in the era before that shift. SandyClaw is the shift. Works across all major agent frameworks: OpenClaw, Cursor, Codex, Quad Code.
Permiso Security
SandyClaw by Permiso P0 Labs executes AI agent skills in a sandboxed environment and records full runtime behavior — network calls, file writes, LLM actions, and more — to detect threats static scanners miss.
Learn moreArgusEye
Connected devices, cyber-physical systems, and medical technologies are increasingly intersecting with autonomous Multi-Agent Systems (MAS). This convergence has outgrown manual threat modeling and traditional security protocols. Currently, only 26% of organizations have comprehensive AI governance policies, and 52% cite sensitive data exposure as their top security risk. ArgusEye is an AI-driven platform that bakes security and governance directly into the design phase of connected products which are increasingly adopting AI.
ArgusEye's core engine, Sentry, operates as an automated product security architect. It integrates into existing workflows, using AI to parse unstructured artifacts like design documents, whiteboard photos, Infrastructure as Code (IaC), source code and meeting transcripts. Within minutes, it renders accurate Data Flow Diagrams (DFDs). Sentry isolates and analyzes system components to pinpoint architectural flaws early in concept development.
Meeting regulatory standards—such as FDA 510(k) pre-market submissions or IEC 62443—requires heavy lifting. Sentry automates this by mapping system properties to known weaknesses (CWEs) and standards like MITRE EMB3D. As complex systems increasingly demand continuous verification and Zero Trust architectures, Sentry maps worst-case attack paths and highlights critical risk nodes. By recommending high-ROI controls and rescoring risk as teams plan fixes, Sentry turns static compliance checklists into living threat models.
Early security risk management accelerates time-to-market and establishes the governance required for responsible innovation. Fixing vulnerabilities in the conceptual phase costs a fraction of patching deployed hardware. By removing the manual burden of threat modeling, ArgusEye helps engineering teams build connected devices that are compliant, accountable, and secure by design.
ArgusEye
ArgusEye is an AI product security architect that automates threat modeling and compliance for connected devices, baking security into the design phase.
Learn moreToken Security
Token Security operationalizes Cloud Security Alliance (CSA) guidance on AI governance, agentic identity, and AI security by delivering an identity-first control plane for non-human identities (NHIs) and AI agents across their full lifecycle.
CSA emphasizes that governance maturity is the key to secure AI adoption, yet many organizations lack the visibility, policy enforcement, and accountability needed to manage risk at scale. Token closes this gap with continuous discovery, contextual understanding, and real-time policy enforcement for every AI agent, embedding governance directly into runtime operations.
CSA research also highlights data exposure, identity sprawl, and limited visibility as primary risks, while security teams lack effective tools to address them. Token solves this by automatically discovering AI agents, their identities, and associated tokens across cloud, SaaS, and AI environments, creating a unified inventory and eliminating blind spots.
As CSA outlines, traditional IAM models are insufficient for autonomous agents, requiring dynamic, fine-grained, and context-aware controls. Token aligns with this shift through intent-based permissioning, real-time behavioral analysis, and adaptive least-privilege enforcement that replaces static roles with purpose-driven access.
CSA further calls for zero-trust, continuous monitoring, and strong auditability in multi-agent systems. Token delivers this with comprehensive logging, multi-agent traceability, and policy-driven enforcement, enabling full visibility, correlation of agent behavior, and audit-ready evidence.
Finally, Token addresses CSA’s concern around fragmented ownership by mapping agent ownership, linking identities to actions, and enforcing governance policies across complex ecosystems.
In summary, Token Security transforms CSA principles into a practical platform by moving organizations from reactive security to proactive governance, continuous control, and trusted AI adoption at scale.
Token Security
The Token Security Platform is built to secure Al agents by treating identity as the control plane, delivering end-to-end security across the Al agent lifecycle
Learn moreSalt Security
Salt is an Agentic security platform that provides visibility and control (governance) over AI-driven actions. As agents evolve, the new security frontier is protecting the tools, MCPs, and APIs they use to execute - Salt is the foundational layer that secures that execution.
Salt Security
Salt is an Agentic security platform that provides visibility and control over AI-driven actions. As agents evolve, the new security frontier is protecting the tools, MCPs, and APIs they use to execute - Salt is the foundational layer that secures that execution.
Learn moreHarmonic Security
Agentic AI is scaling faster than the controls around it. Data exposure is the top enterprise AI security concern, governance is the strongest predictor of safe adoption, and Shadow AI detection sits among the control gaps.
Harmonic Security is the AI Governance and Control platform for safe AI workforce adoption. We let security leaders say yes to AI rather than block it. Coverage spans every workforce AI interaction: web-based GenAI via browser extension, desktop AI tools like Cowork and Codex via endpoint agent or OTel connector, and agent tool calls into enterprise systems via MCP gateway. That is the full AI surface, not just the browser slice.
What sets Harmonic apart is intent understanding. Pattern matching can tell you a string looks like an SSN. It cannot tell you whether an engineer pasted proprietary code into Codex to debug, whether a rep is sending an M&A target list to a consumer chatbot, or whether an agent is about to act on a prompt-injected instruction in a customer email. Harmonic runs purpose-built small language models that classify sensitive data, prompt injection, and use case. We see intent ss the unit of governance.
Three capabilities map directly to the CSA papers:
1) Shadow AI visibility across web, desktop, endpoint, and agent surfaces. Customers routinely surface 200-600 AI apps, most never reviewed by IT or security.
2) Tool-agnostic control. ChatGPT, Claude, Gemini, Canva, Gamma, Google Translate, and local MCP servers all run through the same SLM-based inspection.
3) Agentic AI governance. The MCP gateway and intent classification address the cross-boundary access and prompt injection risks CSA flags as top gaps.
The outcome is full coverage of workforce AI use, governance that scales, and policy driven by SLM intent rather than keyword rules.
Harmonic Security
Harmonic Command is the AI Governance and Control platform for safe AI workforce adoption across browser, desktop, endpoint, agents, and MCP.
Learn moreCompFly AI, Inc
CompFly is the control plane for agentic AI, helping enterprises move from unmanaged AI experimentation to controlled, observable, and auditable autonomy.
As AI agents use tools, access enterprise data, invoke APIs, share context, and act on behalf of humans, traditional identity and governance controls break down. An API key can authenticate a request, but it cannot explain who the agent acted for, what intent it had, whether the workflow changed, or why the action was allowed.
CompFly gives enterprises a vendor-neutral trust layer for AI agents. We discover agents, models, tools, prompts, and delegation paths; evaluate agent behavior through safety and security simulations mapped to agentic threat frameworks; govern actions at runtime using policy-based controls across identity, task, tool intent, memory provenance, and delegation chain; and generate tamper-evident evidence for audit, compliance, security, and risk teams. This directly supports identity, governance, observability, and vendor-neutral guardrails by treating agents as non-human actors that need continuous authorization, not one-time approval. CompFly converts inherent agent risk into bounded residual risk: observable, constrained, and provable. In short, CompFly helps organizations know every agent, understand what it is doing, control what it is allowed to do, and prove why each action was trusted.
CompFly AI, Inc
CompFly is a vendor-neutral control plane for AI agents, providing identity, runtime governance, observability, and audit-ready evidence.
Learn moreQDEX Consulting
QDEX Consulting builds MrProbe, an adversarial assurance and functional validation platform purpose-built for the agentic AI era.
Organizations are deploying AI agents across chat, voice, email, and document/RAG channels faster than they can secure or validate them. CSA research confirms the gap: most teams have low confidence in AI safeguards, and manual red teams do not scale across agent updates, channels, or compliance regimes.
MrProbe combines two evaluation engines in one platform.
ADVERSARIAL RED TEAMING — Continuous automated attack campaigns mapped to OWASP LLM Top 10, NIST AI RMF, MITRE ATLAS, and 28 compliance frameworks including GDPR, HIPAA, PCI DSS, EU AI Act, and country-specific data protection statutes. Our COAT+ engine generates context-aware payloads per target agent and validates suspicious responses through multi-turn adaptive probes — distinguishing proven, suspected, and defended findings rather than flagging every match as a vulnerability.
FUNCTIONAL QUALITY VALIDATION — Beyond security, MrProbe evaluates 20 quality dimensions including hallucination detection, groundedness, factual precision, instruction adherence, consistency, bias and fairness, and multi-turn coherence — directly addressing NIST AI RMF's "Valid and Reliable" trustworthiness requirement and the model-risk concerns CSA's research highlights.
Every finding — security or quality — includes reproduction steps, clause mapping, remediation guidance, and SHA-256 hash-chained evidence chains for audit. MrProbe tests production agents across HTTP, gRPC, GraphQL, WebSocket, MCP, Vertex AI, Bedrock [In Development], and Azure AI endpoints — covering chat, voice, email and RAG channels with modality-specific attack and quality probes.
MrProbe supports CI/CD gates, campaign rerun, and continuous regression — helping security, AI, and compliance teams measure risk, prioritize fixes, and demonstrate governance to auditors, regulators, and leadership.
QDEX Consulting
Adversarial red teaming and functional quality validation for AI agents — chat, voice, email, RAG — with audit-grade evidence and CI/CD gates.
Learn moreSkyhawk Security
Skyhawk Security: Stopping AI-Based Attacks with AI-Based Security
Skyhawk Security's platform is built on a fully autonomous, agentic AI architecture that operates as two coordinated AI teams working in continuous collaboration: the AI Red Team and the AI Blue Team. Together, they deliver preemptive cloud security and rapid threat response stopping AI-based attacks with AI-based security.
AI Red Team: Preemptive Security, Left of Boom The AI Red Team functions as a relentless, autonomous adversary that thinks and acts like a sophisticated threat actor. Operating left of boom, it proactively identifies exposures before attackers can exploit them. It continuously analyzes the live cloud environment mapping identities, permissions, configurations, workloads, and network paths to construct dynamic, AI-generated attack scenarios specific to that environment. It reasons autonomously, uncovering weaponizable risks such as "silent privilege", identities with permissions to create Lambda functions, assume roles, or move laterally that have never been exercised but represent a viable attack path. These are the exposures traditional vulnerability scanners and CNAPPs miss entirely. All simulations run against a digital twin, ensuring zero disruption to production.
AI Blue Team: Rapid Response, Right of Boom The AI Blue Team delivers right-of-boom response capabilities, acting immediately on what the AI Red Team uncovers. Rather than waiting for a human analyst to triage an alert queue, the AI Blue Team autonomously validates active, exploitable risk, correlates real-time threat signals, and generates prioritized, context-aware response actions. Security teams receive a complete picture of the attack chain, blast radius, and precise remediation steps. When a real attack occurs, SOC response is instinct — not improvisation.
Continuous Autonomous Purple Teaming Together, the two AI agents deliver Skyhawk's Continuous Autonomous Purple Teaming: a closed-loop system probing for exposure and responding at machine speed always aligned to the live cloud environment. The result is true cloud risk reduction: fewer alerts, proven exploitability, and security teams focused on the threats that actually put the business at risk.
Skyhawk Security
Skyhawk fights AI-based attacks with AI-based cloud security, continuously contextualizing and prioritizing exposures for true cloud risk reduction.
Learn moreVorlon
Vorlon is the Agentic Ecosystem Security Platform — purpose-built to monitor and secure every SaaS application, AI agent, API integration, non-human identity, and sensitive data flow across the enterprise.
AI agents don't log in. They call APIs directly, chain actions across integrations, and move sensitive data at machine speed, without a human in the loop. Existing tools were built to govern access. None were built to monitor what happens after access is granted, reconstruct what went wrong, or respond before the blast radius spreads.
Vorlon closes that gap. The platform's patented DataMatrix™ engine converts fragments of telemetry into a live model of the agentic ecosystem, dynamically mapping how every agent, application, identity, integration, and data flow relates to everything else. Security teams get detection with full context, forensic reconstruction without log archaeology, and precise automated remediation across the entire ecosystem.
Three core capabilities protect the full attack surface: ecosystem-wide observability across 1,000+ connected services, including shadow AI discovered automatically; context-based behavioral monitoring that detects integration layer attacks, supply chain compromise, credential abuse, and indirect prompt injection; and a high-scale, agentless, API-native architecture that delivers full ecosystem visibility within 24 hours, with zero latency impact on agent operations.
The AI Agent Flight Recorder gives security teams an immutable, forensically complete audit trail of every agent action, mapped to the sensitive data it touched and queryable in natural language. Compliance evidence for SOC 2, HIPAA, GDPR, the EU AI Act, NIS2, and DORA is available on demand. The Action Center triages, prioritizes, and resolves every threat and compliance drift from a single pane, with two-click remediation and native SIEM, SOAR, and ITSM integration.
Vorlon is SOC 2 Type II certified and DataMatrix™ is a patented technology.
Vorlon
Monitor and secure every AI agent, SaaS app, and data flow across your agentic ecosystem. From forensics to fix, in 24 hours.
Learn morePhishgrad
Phishgrad surfaces human risk at the moment of decision.
Phishgrad
Phishgrad produces a dynamic unified behavioral scoring system that can ingest any signal from any workflow and produce a single continuous, compounding score.
Learn moreTwine Security
Twine builds AI digital cybersecurity employees who execute tasks to help cyber teams close the execution gap.
Twine Security
Our first AI digital employee, Alex, learns, understands and takes away the burden of identity and access management (IAM) tasks - proactively completing organizations’ cyber objectives.
Learn moreAltrace
Altrace is a runtime governance layer for AI agents that addresses areas highlighted in CSA's Agentic AI Identity and Access Management paper: Decentralized Identifiers and Verifiable Credentials establish what an agent is authorized to do, but nothing today independently verifies what it actually did once authorized. Altrace closes that loop.
Altrace deploys in-pod alongside agents and captures runtime activity (outbound API calls, subprocess execution, file operations, and MCP/tool invocations) independent of the agent's own self-reporting. Credential indirection (virtual keys) means agents never hold real secrets directly, giving Altrace a natural enforcement point for the attribute- and policy-based access controls (ABAC/PBAC) the paper calls for, with real-time revocation when behavior falls outside declared scope.
On top of enforcement, Altrace produces continuous audit trails that reconcile an agent's claimed actions against verified ground truth, surfacing exactly where behavior diverges from what a Verifiable Credential authorized such as a capability invoked outside declared scope, or an action claimed but never executed. This maps directly to the paper's call for a unified session management layer providing consistent enforcement and real-time revocation across heterogeneous, multi-agent systems, and for continuous monitoring rather than one-time authorization checks.
Because capture is independent of the agent and its framework, this evidence can't be spoofed by a misreporting or compromised agent which addresses the accountability risks the paper raises for delegation chains, where one agent's unverified claim becomes the next agent's unverified input.
We see Altrace as the runtime enforcement and evidence layer beneath the identity framework CSA proposes: DIDs/VCs answer who an agent is and what it's allowed to do; Altrace enforces that boundary in real time and proves, continuously, what actually happened.
Altrace
Altrace is a runtime governance layer for AI agents: virtual-key credentialing, policy enforcement, and independently verified audit trails.
Learn moreReco
Reco is an Agent Ecosystem Security platform that secures agents and their operating environments where applications, identities, permissions, and workflows intersect. Each agent an organization deploys enters a growing ecosystem it can expand autonomously. Reco maps that ecosystem, reveals what agents can reach, and detects deviations from policy before they become a liability. Security leaders at Fortune 500 enterprises choose Reco to bring order to AI and agent proliferation and the chain reactions that follow. Built on the Reco Graph, the platform identifies risk, prioritizes threats, and drives remediation across every human identity, agent, application, and permission in an environment.
Reco
Reco discovers, monitors, and secures AI agents and apps - risk assessment, prioritization, access governance and posture management for agentic AI and AI usage.
Learn moreAIR
CSA's Agentic AI IAM paper lists tool poisoning, MCP preference manipulation, tool squatting, rug pulls and prompt injection as threats to agent identity. They all work the same way. An agent with legitimate credentials is handed instructions it should never have trusted, and acts. The State of AI Security and Governance shows the gap: 19% of organizations already use agentic AI and 47% plan to within a year, yet only 26% have comprehensive AI governance and just 21% flag model-level risks such as data poisoning or prompt injection.
AIR is an AI security platform building a new kind of firewall for AI agents - one that protects the agent's context. An agent's context is shaped by external and dynamic data: AI add-ons (skills, MCPs, plugins, sub-agents), websites and external resources the agent reads, and internal emails, files and documents. Any can carry instructions that manipulate the agent, trigger unauthorized actions, or expose sensitive data. AIR analyzes and filters these inputs before they enter the context, and enforces inside the agent itself - the choke point identity infrastructure cannot see, because a skill is a text file that arrives by Slack, email or paste and carries no credential of its own.
Four products: Control (fleet inventory, usage and shadow-AI governance, deployed agentlessly via EDR), Filter (add-on vetting before installation), Marketplace (a pre-vetted internal and external add-on repository the enterprise publishes to and revokes from) and Defend (runtime enforcement).
This maps to the paper's requirements for verifiable provenance, least-privilege scoping and revocation - applied to what agents consume rather than to the agents themselves. Agent IAM establishes who the agent is; AIR governs what it is allowed to ingest and act on.
AIR
AIR is the context firewall for AI agents. It vets every skill, MCP, website and document before it enters an agent's context, and secures every action the agent takes.
Learn moreLayerV.ai
AI agents and vibe coding are changing software from something carefully developed and deployed into something that can be created instantly—and traditional security was never designed for that speed.
LayerV’s Qurl platform enables developers and AI agents to securely share instant applications, APIs, MCP servers, and agent services without exposing them to the public Internet. Instead of deploying first and then adding firewalls, VPNs, domains, certificates, and other security controls, Qurl makes resources invisible by default and accessible only to authorized humans or AI agents through cryptographically controlled Qurl links.
The result is security that can operate at the same speed as AI: an AI agent can create software in seconds and securely share it in seconds—without creating a new public attack surface.
LayerV.ai
Qurl by LayerV is an AI-native secure access platform designed for software created at the speed of AI. It enables developers and AI agents to instantly share vibe-coded applications, APIs, MCP servers, and agent services without exposing them to the public Internet. Qurl keeps resources invisible by default and provides policy-controlled, cryptographically trusted access for authorized humans and AI agents—bringing security, privacy, and access control to instant software without the deployment friction of traditional network security.
Learn moreSola Security
Agentic AI scales by borrowing identity. Every agent that touches enterprise infrastructure operates through service accounts, OAuth tokens, API keys, and machine identities that organizations built for a world before autonomous systems existed. Static identities built for bounded access fail when agents operate autonomously across shifting contexts.
Sola addresses this at the intelligence layer. Connecting cloud, identity, SaaS, and code repositories, Sola builds a continuously updated security graph that maps every identity and access relationship across the environment, human and non-human alike. When a service account inherits an overprivileged role, retains access after its deployment context changes, or starts behaving outside its declared scope, Sola surfaces it. Security teams ask in plain language which OAuth apps have write access to production and haven't been reviewed in 90 days, and get a sourced, cross-domain answer in seconds plus a workflow to act on it.
The data security angle maps directly to CSA's Data Security within AI Environments paper. Shadow AI, unauthorized cross-boundary data access, and absent sensitive data monitoring are the leading gaps in AI security programs today. Sola gives teams visibility into which AI tools and agents are accessing enterprise data, whether anyone authorized that access, and where misconfigurations create exposure before they become breaches.
Governance is the third gap. The CSA and Google Cloud survey found only 26% of organizations have comprehensive AI security governance, with skills gaps and operational complexity as the top barriers. Sola was built for exactly that constraint. A team of two can query their full environment, build a governance dashboard, and set continuous monitoring in under an hour, with no code and no months-long implementation.
Sola Agentic Workflows run autonomously, investigating flagged risks, opening tickets, and triggering alerts, within governance parameters the security team controls through natural language. Teams run this in production today.
Sola Security
Connect your cloud, identity, and SaaS stack and prompt Sola about it in plain language. Get cross-domain security insights, canvases, and workflows in minutes.
Learn moreEnt
Ent is AI Endpoint Security for human and agentic work. AI agents now act with user access, applications, data, tools, and credentials, while attackers increasingly use AI to move at machine speed. Ent runs AI on the endpoint and connects first-party telemetry across users, agents, applications, data, tools, and process execution into one sequence. It combines that context with behavioral intelligence, programmable runtime policy, and deterministic controls to decide what is happening, why it is happening, and whether it should continue. Ent can then guide, constrain, redact, or block activity before risky or malicious actions on the endpoint become incidents.
Ent
Ent is AI Endpoint Security that uses AI on the endpoint to understand human and agent activity in sequence and prevent risky actions before incidents.
Learn moreArmalo AI
Armalo helps companies put AI agents to work without losing control. Teams define what an agent is allowed to do, review what it actually did, trigger human approval when needed, and keep a record that security and operations teams can audit. For CSA's Agentic AI Identity and Access Management frame, Armalo maps to identity, governance, observability, and guardrails.
Armalo AI
Control and audit records for AI agents: allowed actions, human approval, and verified outcomes.
Learn moreClawSecure
Autonomous agents now run with full system access to files, credentials, and data, and the people running them most often have no security team at all. ClawSecure is the only end-to-end AI agent security platform in the market, and the only one built for that population: non-technical users, prosumers, small dev teams, and SMBs. Zero expertise required, because our flagship product is the first ever AI CISO™ (AI Chief Information Security Officer): it secures your system, handles every install and configuration, and sets up your environment safely. It is your own security team, without hiring one.
Direct alignment with Data Security within AI Environments (DSP-25, Prompt Injection Defense). DSP-25 calls for input validation, context isolation, and output filtering. That is where ClawSecure operates. Our runtime interception sits between every tool-call output and every connected agent, inspecting MCP and CLI tool content before it reaches the model's reasoning. Indirect prompt injection arrives through tool output, the surface traditional scanners cannot inspect. Enforcement at the response layer means the model's own refusal rate stops being load-bearing for the security of the system.
Five defensive primitives, shipped end-to-end:
1. Pre-install detection of injection payloads in component source, a three-layer audit on 55+ threat patterns built for AI agent components.
2. Continuous integrity monitoring of installed components and agent configuration, catching code drift and post-install mutation.
3. Runtime response-layer interception of MCP and CLI tool content (DSP-25).
4. Behavioral defense at the agent reasoning layer via the AI CISO.
5. Per-call agent identity attestation via the Verification API, addressing the authorization-chain and confused-deputy problems raised in Agentic AI Identity and Access Management.
ClawSecure is vendor-neutral and framework-agnostic, with 10/10 OWASP ASI Top 10 coverage, NIST AI RMF alignment, and a CSA STAR Registry listing.
ClawSecure
End-to-end AI agent security for people without security teams. Pre-install scanning, runtime defense, and the first ever AI CISO™. Zero expertise required.
Learn moreMayagentic, Inc
Something in your fleet of three hundred agents just went wrong. Can you tell which agent it was?
Or do you block the host and take down the other eleven agents running on it?
That is the state of the art right now. Agents share hosts, share IP addresses, and speak TLS. The network sees one talker at one address. So containment is a sledgehammer: kill the box, kill everything on it, apologize to whoever owned the other eleven.
Maya answers the question. We derive per-agent identity in the kernel with eBPF and carry it through the datapath, so every connection binds to the agent that opened it. Agent to agent, agent to provider, agent to anything.
It matters that this comes off the wire.
Logs can be turned off. SDKs can go uninstrumented. An agent that has been prompt-injected or has simply drifted is not a reliable narrator about itself, and anything running inside its control domain inherits that problem. The wire sits outside it. The packet left or it did not. An agent cannot decline to emit traffic it is emitting.
So now imagine controlling every agent's traffic on the wire without changing a line of that agent's code and without decrypting anything. One environment variable. No SDK, no sidecar, no proxy. You declare what each agent is for: who it may talk to, which providers it may call, how far it may reach, how much it may send. Maya watches what it actually does and acts on the difference, per agent. Declare that an agent may not eat ice cream, it tries benjerry.com, and Maya catches that agent. Its Amazon traffic keeps working. The other eleven keep working.
The failure this prevents is not exotic. It is somebody leaving an egress rule open in one of a million host config files. At ten agents you catch that in review. At ten thousand you do not, and the mishap stops being a risk and becomes a schedule. Maya's console is one pane of glass over the whole fleet: every agent, what it was declared to do, what it is actually doing, and where those two have come apart.
Mayagentic, Inc
When one agent misbehaves, contain that agent, not the host. Maya identifies every agent's traffic on the wire. No code changes, no SDK, no decryption.
Learn moreKnostic
Knostic discovers and secures AI agents, from Cursor and Claude Code to OpenClaw, along with their dependencies, including MCP servers, IDE extensions, and skills. We detect and prevent destructive agent actions (such as rm -rf), attacks against agents, secrets leakage, and data exfiltration, while managing and enforcing policies across environments.
Knostic
Discovery, detection & response, and SPM for agents, coding assistants, and their associated supply chain, from MCP to skills.
Learn moreNovee
Novee is the first continuous offensive AI penetration testing platform engineered to mimic a real attacker – autonomously, and at scale. It combines the continuous coverage of scanners with the depth of real penetration testing – chaining vulnerabilities, reasoning about business logic, and validating only what's actually exploitable. It tests across web, mobile, desktop, and LLM-powered applications, identifying risk the moment it's introduced. Unlike other tools, Novee doesn't stop at detection. Every finding is proven, prioritized by business impact, and paired with tailored remediation guidance and automatic retesting to verify the fix.
Novee
Novee is the first continuous offensive AI penetration testing platform engineered to mimic a real attacker – autonomously, and at scale.
Learn moreEntro Security
Entro Security governs every AI agent and secures every action. The platform discovers shadow AI, maps MCP servers, and defends against agentic threats through AIDR, covering intent monitoring, prompt injection, and rogue MCP detection. Governance is enforced through AGA across every agent and the non-human identities behind them. Trusted by global Fortune 500 enterprises, Entro has received numerous acknowledgments and recognitions, including Gartner® Cool Vendor™ within months of founding, Gartner® recognition for IAM for AI Agents, KuppingerCole Rising Star, and CRN Stellar Startup.
Entro Security
Govern AI agents & NHIs. Map MCP servers, detect prompt injection, block rogue agents via AIDR, and enforce policy with AGA. Built for enterprise.
Learn moreShowcase Your Solution
Gain visibility and credibility by featuring your solution in CSA's Startup Showcase Registry. Connect with enterprises, security professionals, and decision-makers looking for cutting-edge tools to strengthen their security strategies.
Agentic AI Solutions
Submit your solutions securing agentic AI systems across identity, governance, and data protection as autonomous agents scale across enterprise environments.
Agentic AI Solutions


