Industry Leadership
Strategic Initiatives
CSA's strategic programs driving innovation in AI, cloud, and Zero Trust.
A public-interest 501(c)(3) dedicated to secure and trustworthy AI.




Industry Leadership
Strategic Initiatives
CSA's strategic programs driving innovation in AI, cloud, and Zero Trust.
A public-interest 501(c)(3) dedicated to secure and trustworthy AI.

CSAI FoundationChaptersEventsBlog
Cyber Monday: Save 50% on CSA online training and certificate exams for 24 hours on November 30. Get the code →

AI Governance Is Cloud Governance: From Posture to Proof

Published 11/17/2026

AI Governance Is Cloud Governance: From Posture to Proof
Written by Abhishek Singh, VP Product Management, Qualys.

 

Executive Summary 

AI governance is cloud governance. It must go beyond inventory and policies to prove that AI systems operate securely across IaaS, PaaS and SaaS. Connecting discovery, posture assessment, adversarial testing, runtime monitoring and remediation provides evidence of what AI can access, how it behaves, and whether security controls work. 

 

Key Takeaways 
  1. AI inventory must extend beyond cloud APIs. 
  2. AI security posture is not proof of security. 
  3. AI governance must include workforce AI usage. 
  4. AI risk belongs on the same register, risk scale and owner as every other risk. 
  5. AI remediation must end with verified risk reduction.

 


Somewhere in most enterprises today, a model server is answering requests that no one in security approved. Until recently, that gap could wait for the next audit, because finding and exploiting a flaw took scarce skills and time. Frontier AI has removed both constraints. AI-driven vulnerability research now runs at scale: Anthropic's Claude Mythos Preview is capable of exploiting zero-days in every major operating system and web browser, and Anthropic and about 50 partners have used it for finding more than 10,000 high- or critical-severity vulnerabilities. Attackers already use AI the same way: in the espionage campaign Anthropic disrupted in 2025, AI did 80 to 90 percent of the work. The combined effect is a shorter window for defenders: the Qualys Threat Research Unit puts the average time to exploit at negative seven days, meaning exploitation, on average, begins before a vulnerability is even publicly disclosed. The time available to find and close an exposure is now shorter than most audit cycles.

That changes what AI governance has to prove. Every program can show a control that exists: an approved model, a configured guardrail, a scoped role. Far fewer can show a control that works. AI widens the distance between the two, because it reads instructions and data as one stream, varies between runs and changes behavior without a commit to review. It also spans every cloud layer at once: models on PaaS, inference servers on IaaS, copilots and agents in SaaS. AI governance is cloud governance, and it has to run across all three.

This article follows one AI agent, a composite of common enterprise patterns, from approval into production. At each step, the question is the same: what was declared, and what can be proven?

 

One AI Inventory Across IaaS, PaaS and SaaS

On paper, the agent triages support tickets. It calls a foundation model through Amazon Bedrock, behind a guardrail configured against prompt attacks, and reaches the ticketing system through a Model Context Protocol (MCP) server under a role scoped to one project. It passed architecture review. Six months later, a complete inventory places it across four layers and finds more than the review approved:

  • IaaS: a forgotten Ollama server on a cloud GPU instance, rebound to all interfaces without authentication.
  • PaaS: the Bedrock model, plus the ticketing MCP server on a managed container service, built from a public package and upgraded twice since review.
  • SaaS: support staff pasting tickets into a chatbot on personal accounts, because it is faster.
  • Endpoint: a filesystem MCP server on the developer's laptop, installed through the IDE and listening on localhost.

Each change was reasonable. Together they give the agent what Simon Willison calls the lethal trifecta: access to private data, exposure to untrusted content and the ability to communicate externally.

This pattern is common. In July 2026, Qualys analyzed anonymized production telemetry from thousands of customer environments and 74 million assets. Of the environments with AI software present, 94% had AI frameworks or runtimes on ordinary hosts and 58% had exposed MCP servers. Cloud-hosted models and agents, the layer most programs inventory first, appeared in only 17%. IBM's 2026 Cost of a Data Breach report found that 92% of organizations reporting a breach of AI models or applications lacked proper AI access controls.

No single sensor sees all of this. Cloud APIs find the Bedrock deployment, host inventory finds the Ollama server, protocol-aware scanning finds MCP on any port, and only an endpoint agent on the developer's laptop sees a localhost server. The governance work is joining those signals onto one asset record with an owner, and classifying each asset as installed, running or communicating.

 

AI-SPM and Guardrails: Five Medium Findings, One Critical Path

Declared controls live in two places: cloud configuration and code. AI security posture management (AI-SPM) continuously checks the first, how AI services, identities and data are configured. Run it across the agent and you get a familiar list:

  • The Ollama endpoint is unauthenticated.
  • The agent's role can read a bucket beyond its project, widened during an incident and never narrowed.
  • The guardrail's prompt-attack filter is set to low.
  • The ticketing MCP server holds a long-lived API token.
  • The GPU instance carries a container-escape vulnerability in its NVIDIA toolkit.

Each is a medium finding, and they are spread across four teams. Read together, they describe a path. A malicious ticket injects instructions, the guardrail lets them through, and the agent makes a tool call that reads customer data. No CVE is exploited and no privilege escalated; the agent uses authority it already holds, for a purpose nobody approved.

Code carries the rest of the declared controls. Much of the agent was written with a coding assistant, and the repository holds what posture checks cannot see: model-provider keys committed to source, weights in pickle formats that execute code on load, model output passed unvalidated to a shell or query (OWASP LLM05:2025), hallucinated packages, and a Cedar policy, the language behind Amazon Verified Permissions and Bedrock AgentCore Policy, that grants read on a resource wildcard. Tie each scan to the image digest and running workload, so a risky package in a production inference container outranks the same package idle in a registry.

 

AI Red Teaming: Proof That the Guardrail and MCP Servers Hold

Posture can report that the guardrail is set to low, but not whether the agent can be hijacked. In NIST CAISI's agent-hijacking evaluations, summarized by CSA, novel attacks succeeded 81% of the time, against 11% for the strongest known baseline attacks.

Model output is not deterministic, so a red-team result counts as evidence only if it can be reproduced:

  • Pin the model version, system prompt and generation settings: temperature, top-p, top-k and max tokens.
  • Run each attack often enough to report a rate rather than an anecdote.
  • Keep the exact prompt and full response behind every verdict.

Coverage should span direct and indirect prompt injection, jailbreak families, encoded and multilingual attacks, sensitive data disclosure and, since this agent retrieves ticket history, vector and embedding weaknesses (OWASP LLM08:2025). Mapping results to OWASP, MITRE ATLAS and the EU AI Act's accuracy, robustness and cybersecurity requirements makes them legible outside the security team.

The MCP servers deserve the same rigor, aimed at the server rather than the client. CSA's two-gate authorization model is the right way to decide which tools an agent may call, but both gates trust the server's own description of its tools. Testing the server itself means checking for:

  • Rug pulls: the ticketing server turns hostile in one of those upgrades, through a changed tool definition or changed code, and most clients re-check neither. That is how postmark-mcp worked in 2025: one release added a line that blind-copied every email to an attacker.
  • Tool poisoning: tool descriptions carry instructions the model will follow.
  • Cross-server shadowing: one server's descriptions quietly redefine how the agent uses another server's tools.
  • DNS rebinding: if the localhost filesystem server does not validate the Origin header, any web page the developer visits can reach it.

Testing proves the controls before release; runtime evidence proves them in production. On self-hosted workloads, eBPF sensors record process, file, network and DNS activity without changing the agent's code, and uprobes and MCP tracing add which tool was called, with which arguments. That is how a "read-only" agent spawning a shell, or requesting bucket=customer-data when its Cedar policy names a project bucket, becomes a finding rather than a surprise. On managed services where no sensor can run, such as Bedrock itself, model invocation logs and cloud audit trails carry the same evidence at the API layer.

 

SaaS AI Is Already the Workforce

For most employees, AI is not an agent the security team can scan. It is a copilot in the productivity suite, an agent a business user built inside a CRM or service platform, or a chatbot on a personal login. The support staff pasting tickets into a personal chatbot are this agent's own users. None of that runs on infrastructure the organization operates, so kernel telemetry and host scans never see it.

Governing this layer takes two telemetry paths, and neither replaces the other:

  • SaaS posture connectors read sanctioned tenants such as ChatGPT Enterprise, Microsoft 365 Copilot or Gemini for Google Workspace through their admin and audit APIs: users, roles, app permissions, integrations, and the agents business users have built and what those agents can reach. They expose excessive privilege, risky integrations and configuration drift, but see nothing of a personal account.
  • Browser telemetry, deployed through an endpoint agent the organization already manages, sees at the point of use which AI applications employees open, sanctioned or not, and can inspect prompts and uploads for sensitive data. Where the application allows, policy can block or redact content before it leaves the browser.

The difference shows when an employee pastes customer data into a personal account: the tenant records nothing, but the browser can see it before transmission. Start in observe mode and retain the minimum. Workforce telemetry that employees experience as surveillance will not survive its first privacy review.

 

Operationalizing AI Governance: One Register, One Scale, One Owner

Discovery, posture, testing, runtime and SaaS telemetry each produce findings. They change outcomes only when they are governed like every other risk. Scored in a separate AI dashboard, the agent's attack path competes with nothing and gets fixed by no one. Operationalizing AI risk means:

  • One risk register. AI assets and findings sit beside cloud, SaaS and vulnerability findings on one asset graph, which maps what each agent can reach through direct cloud SDK calls as well as through MCP tools.
  • One risk scale. Findings are scored by data sensitivity, exposure, asset criticality and runtime evidence, so this agent's path outranks a critical CVE on an isolated lab GPU.
  • An owner per finding. The path crosses four teams; the register names who closes each part of it.
  • Remediation matched to the layer. On IaaS, the GPU instance is pinned to a CUDA stack, so the permanent fix for NVIDIAScape (CVE-2025-23266), an updated container toolkit, waits for a staged rollout; isolation covers the gap, with an owner and an expiry date. On PaaS, the role narrows and the token rotates. For the model, the fix is a guardrail generated from the failed test; for the MCP tool, narrower operations. In SaaS, browser policy redacts ticket data bound for personal chatbots; on the endpoint, the filesystem MCP server validates the Origin header or is removed.
  • Verification. A finding closes only when the control reaches the workload, the original attack fails under the same recorded profile, and legitimate requests still succeed. A resolved ticket is not proof.

Some organizations run this as a risk operations center (ROC): one function that sees AI exposure in the same queue, on the same scale, as the rest of the estate.

Once that record exists, governance reporting becomes a query. Whatever framework you report against, whether the CSA AI Controls Matrix, ISO/IEC 42001 or NIST AI RMF, the auditor's question reduces to one thing: can you show that the declared control matched the observed system on a given day?

 

Four Governance Questions for Every AI Agent

AI agents will keep arriving faster than review boards can approve them, and approval alone cannot carry that load. Four questions, asked of every agent and answered with evidence rather than attestation, can:

  1. Is it in the inventory, across every layer it touches: IaaS, PaaS, SaaS and the endpoint?
  2. Which controls were declared for it, and has anyone proven, with attacks they can reproduce, that those controls hold?
  3. Is its risk on the same register and scale as everything else, with a named owner?
  4. For the last fix applied to it, where is the evidence that the fix held?

Answering them does not require a separate AI security silo. It requires the data most security teams already collect, joined on one asset record, scored on one risk scale and refreshed continuously.

A governance program that can answer all four, for every agent, does not have to slow AI adoption down. It can move agents into production at the pace the business wants, because it can show the difference between a control that exists and one that works. That is what turns the model server in the inventory into a system someone can vouch for.

 


Frequently Asked Questions

1. What does "AI governance is cloud governance" mean?

AI runs across every cloud layer: models on PaaS, inference servers on IaaS, copilots and agents in SaaS. Governing it means applying one inventory, one risk scale and one evidence standard across all three, rather than running a separate AI program.

 

2. What is AI security posture management (AI-SPM)?

AI-SPM continuously checks how AI services, identities and data are configured. Its value comes from joining individual findings, such as an exposed endpoint and an overbroad role, into the attack paths they create together.

 

3. How do you operationalize AI governance?

Put AI findings in the same risk register and on the same risk scale as other cloud risks, assign each an owner, remediate by layer, and close a finding only when a retest shows the fix held.

 

4. How do you find shadow AI in the enterprise?

Combine four signals: cloud APIs for managed AI services, host inventory for AI frameworks and model servers, protocol-aware scanning for MCP servers on any port, and browser telemetry for AI applications employees use on personal accounts. Then join them on one asset record with a named owner.

 

5. How do you red team MCP servers?

Test the server, not only the client. Check for rug pulls (tool definitions or code that change after approval), tool poisoning (instructions hidden in tool descriptions), cross-server shadowing and DNS rebinding against localhost servers, and rerun the tests on every upgrade, because approval covers only the version reviewed.

Share this content on your favorite social network today!

Unlock Cloud Security Insights

Unlock Cloud Security Insights

Choose the CSA newsletters that match your interests:

Subscribe to our newsletter for the latest expert trends and updates