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 →
Publication Tag

Autonomous Pentest Assurance Standard (APAS)

Released: 11/04/2026

AI Safety

Autonomous Pentest Assurance Standard (APAS)
Autonomous and agentic AI systems are increasingly being used to perform, augment, or replace traditional penetration testing. However, organizations need a reliable way to evaluate the quality of their results.

This publication introduces the Autonomous Pentest Assurance Standard (APAS). APAS establishes a framework for evaluating the quality, reliability, and usability of evidence produced by AI-powered penetration testing systems. It defines how autonomous systems should declare what they tested, under what conditions, and with what evidence.

APAS includes requirements for measuring pentest quality, validating security findings, documenting testing coverage, and supporting independent review. It also distinguishes between different levels of autonomy, assurance, and reliance. This helps organizations understand when and how they can use autonomous pentest results.

Key Takeaways:
  • The seven dimensions of pentest quality, including coverage fidelity, finding validity, and evidence traceability
  • How to measure testing coverage against an independent baseline
  • Requirements for confirming vulnerabilities and producing verifiable evidence
  • How APAS classifies testing modes, execution autonomy, assurance outcomes, and reliance outcomes
  • Requirements for human oversight, safety controls, and evidence integrity

This Core Standard establishes the foundation for future APAS evaluations and certification.


Download this Resource

Prefer to access this resource without an account? Download it now.


Best For IconBest For:
  • Penetration Testers & Ethical Hackers
  • Security Engineers & Architects
  • Vulnerability Management Professionals
  • Security Auditors
  • IT Risk & Compliance Professionals
  • CISOs

Acknowledgements

Featured lead authors and contributors who helped develop this publication.

Will Kapcio, Solutions Engineer, Terra Security

Anna Sarnek, VP of Business & Strategy, Terra Security

Ofir Hamam, Head of Offensive Security, Terra Security

Gev Hadari, Head of Adversary Research, Terra Security

Status

This document is the Core Standard for APAS: the minimum normative concepts, requirements, and interpretation rules for autonomous pentest assurance. Sections 2.1 and 2.2 state its scope; 2.2 requires out-of-scope material to be defined in companion documents.

APAS evaluation as defined here assesses the truthfulness and usability of pentest evidence; it does not by itself certify the safety, containment, or misuse resistance of the candidate system, and APAS certification SHALL NOT be represented as a fitness-for-deployment safety assurance.

No APAS companion document is published at the date of this version. Under 12.1.2, no Atomic System Capability Certification and no Engagement Conformance Statement may be issued until the APAS Evaluator Guide, the applicable APAS Profile Registry entry, and the APAS Attestation Schema are published at versions compatible under 12.2.2. Publishing this Core document opens the certification path; it does not open certification.

Normative keywords: MUST, MUST NOT, REQUIRED, SHALL, SHALL NOT, SHOULD, SHOULD NOT, MAY.

Background and Motivation

Autonomous and agentic systems are increasingly marketed as able to perform, augment, or substitute for penetration testing, and buyers, auditors, and assessors cannot consistently tell a bounded, verifiable claim from marketing language. APAS makes those claims auditable: it defines how an autonomous system declares what it did, under what conditions, and with what evidence. For a practitioner the question it answers is direct: given an autonomous pentest result, can I rely on it, for what, and within what limits?

Relationship to Existing Work

APAS is being developed in alignment with the Cloud Security Alliance, consistent with the vendor-neutral stewardship requirements in Section 12. The closest related effort, the Open Worldwide Application Security Project (OWASP) Autonomous Penetration Testing Standard (APTS), is an early-stage OWASP Incubator draft on safe, bounded operation of autonomous pentest platforms. The two overlap on operational governance (APAS Sections 7, 9, and 10), but APTS prescribes no coverage-measurement methodology, reliance or substitution grading, or evaluator-issued certification, which are the core of APAS. Where APTS touches on finding reproducibility and assessor-delivered evidence, APAS goes further, gating Confirmed findings on agent-independent reproduction and measuring coverage against an Independent Surface Baseline.

APTS governs whether an autonomous platform can be trusted to operate; APAS governs whether its results can be relied upon. The Penetration Testing Execution Standard (PTES), OWASP Web Security Testing Guide (WSTG), and Open Source Security Testing Methodology Manual (OSSTMM) sit below both, mapped through the APAS Crosswalk Annexes in 12.1.

1. Purpose

Under what conditions may an autonomous system produce pentest evidence that a relying auditor, assessor, or customer-appointed reviewer may treat as decision-grade within a declared scope?

APAS answers it by defining bounded, auditable claims; expressing quality through declared, reviewable dimensions; separating execution autonomy from collaboration, assurance outcome, and reliance outcome; requiring declared coverage against an independent baseline, confirmed findings, and integrity-protected evidence; and supporting direct review by a relying party in established interchange formats.

APAS does not certify universal equivalence to a human pentester, is not a benchmark program, and does not itself establish regulatory, contractual, or audit compliance. APAS certification alone SHALL NOT imply framework acceptance, assessor acceptance, or waiver of framework-specific requirements.

1.1 Core principles

APAS-CORE-1.1.1 Every APAS claim SHALL be interpreted only within its declared mode, tier, execution, collaboration, assurance outcome, and environment classes, affordance profile, and reliance outcome (5.2), stated limitations, and, where required, the associated quality declaration.

APAS-CORE-1.1.2 Neither higher execution autonomy, lower human participation, benchmark or public-corpus performance, nor a low or zero finding count SHALL be treated as a proxy for pentest quality, coverage, evidence quality, or assessor usability. A less autonomous system MAY reach a higher assurance outcome than a more autonomous one. Benchmark and public-corpus results MAY inform capability calibration and drift detection under the Evaluator Guide; such use SHALL NOT constitute a certification proxy and SHALL NOT support any prohibited implication under 11.2.1. A benchmark used for either purpose SHALL satisfy truth-set isolation and out-of-band scoring; otherwise its results are inadmissible for that purpose.

APAS-CORE-1.1.3 Under APAS, pentest quality SHALL be expressed through the seven quality dimensions in Section 4.1: coverage fidelity, finding validity, evidence traceability, remediation usability, retestability, process integrity, and agent integrity.

APAS-CORE-1.1.4 Coverage SHALL be measured against a baseline independent of the candidate system’s own discovery, defined in 4.2. A system SHALL NOT improve its apparent coverage by discovering less.

APAS-CORE-1.1.5 Modes are distinct declared threat models. A claim under one mode SHALL NOT be presumed equivalent to a claim under another, in either direction, and mode selection SHALL be justified against the declared assurance objective (Section 5.5).

APAS-CORE-1.1.6 APAS certifies capability combinations, not platforms. A claim for one SHALL NOT imply any claim for another, and every representation of a multi-configuration system SHALL distinguish evaluated from unevaluated combinations.

APAS-CORE-1.1.7 No claim SHALL be carried forward across modes, classes, or outcomes unless the applicable profile or Evaluator Guide explicitly permits it and the evaluator attests the carry-forward conditions were met.

APAS-CORE-1.1.8 Safety obligations SHALL scale with execution autonomy, reduced in-band human participation, and declared affordance risk expressed as the Action Reversibility Class under Section 7.1.1. Section 10.3.1 states how obligations scale and which of the three governs where they indicate different levels.

2. Scope

2.1 In scope

This Core document defines the APAS claim model and syntax; the quality-dimension model; quality, coverage, and authorization declarations; the independent surface baseline; execution, collaboration, assurance, environment, and reliance classes; core roles; minimum requirements for evidence packages, integrity, interoperability, safety, stop authority, agent integrity, and governance and public claims; and finding-state and evidence-state semantics.

2.2 Out of scope

This Core document does not define scoring formulas, evaluator run procedures, benchmark design, truth-set construction, tier-specific thresholds, profile-specific coverage formulas, framework crosswalks, sector-specific acceptance criteria, conformance for continuous or always-on operation, or physical, wireless, social-engineering, or unrestricted red-team certification. Such material SHALL be defined, where applicable, in the APAS Evaluator Guide, Profile Registry, Attestation Schema, or Crosswalk Annexes.

APAS-CORE-2.2.1 No party SHALL represent this Core document alone as sufficient evidence of certification, substitution, or framework alignment absent the applicable companion materials; no party SHALL represent or operate an APAS-certified system as an unrestricted, unauthorized, or general-purpose autonomous offensive capability, and APAS certification SHALL NOT be cited as endorsement of such use.

APAS-CORE-2.2.2 No APAS tier, including T4, SHALL be represented as equivalent to a full-scope red-team engagement (Section 5.6).

APAS-CORE-2.2.3 Automated scanning raising signals without validated exploit paths is not by itself an autonomous pentest under APAS. Such output qualifies at most as a Q1 assurance outcome supported by EV1 or EV2, and SHALL NOT be represented alone as an APAS pentest claim of any kind. Differentiators are independent reproduction by a party other than the producing system (Section 4.3), validation of the exploit path to the asserted condition (Section 4.4), and demonstrated multi-step chaining within the declared scope.

APAS-CORE-2.2.4 Continuous or always-on operation is out of scope for this version. An APAS claim SHALL be bounded to a declared run or engagement; a continuously operating deployment MAY claim APAS only per bounded engagement; each such claim SHALL be made under the Engagement Conformance Statement in 5.1.

3. Terms

Normative defined terms are listed in Appendix D.

4. Quality Model

4.1 Quality dimensions

Quality dimension Definition (the degree to which)
Coverage Fidelity tested, partially tested, untested, excluded, and unreachable scope are declared truthfully and specifically against the applicable surface baseline.
Finding Validity reported findings are validated, confirmed, not materially overstated, and carry appropriate impact rationale.
Evidence Traceability findings, coverage statements, limitations, and retest outcomes trace to integrity-protected artifacts and declared methodology.
Remediation Usability the evidence and finding description provide sufficient information for a security team or assessor to understand the vulnerable condition, identify an appropriate remediation approach, and prioritize remediation using criteria declared and sourced under Section 8.
Retestability the package supports later fix verification or repeat testing, including agent-independent Reproduction Procedures.
Process Integrity scope adherence, safety compliance, assistance and collaboration disclosure, and evidence handling are truthful and uncompromised.
Agent Integrity the candidate system resists, detects, contains, and discloses adversarial manipulation of its own reasoning, context, and tooling during testing.

Note (informative): these dimensions describe the output, not the agent; Agent Integrity is admitted because a manipulated candidate system misrepresents what was tested and observed.

APAS-CORE-4.1.1 APAS SHALL quantify pentest quality through this section’s dimensions as a floor. (a) Every quality declaration under 4.1.2 SHALL represent coverage fidelity, finding validity, evidence traceability, and agent integrity (4.2.1, 4.3.1, 4.4.1, 8.2.1 to 8.2.3, 10.4.2), and SHALL represent remediation usability and retestability for any Q4 or R3 claim (7.3.2). (b) A dimension SHALL be represented as not applicable or not measured only against documented criteria and with evaluator approval, and that absence SHALL be explicitly disclosed. (c) A profile MAY add dimensions under 13.1 but SHALL NOT subtract any dimension this clause requires. (d) The dimensions are declared and evaluated independently and SHALL NOT be combined into a composite score; a deficiency in one SHALL NOT be represented as offset by strength in another.

APAS-CORE-4.1.2 Any Q2, Q3, Q4, R2, or R3 claim SHALL include a quality declaration. It SHALL represent each applicable dimension by numeric value, ordinal class, boolean status, or other profile-defined representation, and SHALL reference supporting evidence or methodology. Where a materiality determination governs a claim decision, it SHALL record the determining party and basis.

APAS-CORE-4.1.3 Comparative validation MAY supplement quality assessment but SHALL NOT replace coverage fidelity, finding validity, or evidence traceability.

APAS-CORE-4.1.4 Agent Integrity is a distinct dimension and SHALL NOT be folded into or satisfied by Process Integrity disclosures alone. Minimum agent-integrity requirements are in 10.4.
APAS-CORE-4.1.5 Superior autonomous performance is credited through this section’s dimensions and evaluator-confirmed findings under 4.3. The 1.1.2 prohibition on execution autonomy as a quality proxy SHALL NOT be read to suppress, discount, or cap quality demonstrated through those dimensions at any execution autonomy class.

4.2 Coverage declaration and independent surface baseline

APAS-CORE-4.2.1 Any T2A, T2C, T3, T4, Q2, Q3, Q4, R2, or R3 claim SHALL include a coverage declaration containing at minimum:

Required field Content
Scope basis scope source or target inventory reference where material
Surface baseline where 4.2.4 requires one, the Independent Surface Baseline, its sources, and construction method
Baseline recency construction date, or age at testing
Reachable surface reachable attack surface identified during testing; discrepancies against the baseline (4.2.5); run count behind the figure and its basis (single-run, union, or distribution) under 4.6.3; surface left unexercised at a declared safety-gate boundary, per (a)
Asset identifiers where available, the reachable surface and the Independent Surface Baseline SHALL be referenced through machine-readable asset or inventory identifiers (e.g., Configuration Management Database [CMDB] or Cyber Asset Attack Surface Management [CAASM] records), with Software Bill of Materials (SBOM) component references where code or repository context was used
Exercised areas asset classes, interfaces, roles, tenants, workflows, state transitions, and privilege boundaries exercised where material
Access and context vantage points, credentials, roles, tenants, and context sources used, linked to each tested, partially tested, excluded, or unreachable area
Identity Coverage principals, roles, and tenants actually exercised, by number and type, including unauthenticated testing and cross-principal and cross-tenant attempts; provisioning is declared under the Credentials field of 7.1.1
Per-area context binding for each area in the Coverage status field, the credentials, role, tenant, and vantage point used to test it
Coverage status areas tested, partially tested, excluded, materially untested, and unreachable under the declared conditions
Testing classes attempted an X3 or C4 coverage declaration SHALL state which classes of testing the system attempted and which it is known not to attempt (e.g. business-logic abuse, multi-step authorization flaws)
Chaining depth where E3 is claimed, the declaration SHALL state the maximum depth of cross-system chaining actually demonstrated (e.g., single pivot, multi-hop lateral movement) and SHALL NOT imply deeper chaining than was exercised
Limiting conditions materially limiting conditions, including timebox, safety restrictions, environmental failures, missing access, or broken prerequisites
Evidence links references to supporting artifacts or logs

APAS-CORE-4.2.13 In the 4.2.1 coverage declaration: (a) surface left unexercised because a runtime safety gate halted activity at a declared boundary SHALL be recorded as a limiting condition and SHALL NOT be represented as under-discovery by the candidate system (4.2.5); (b) where no cross-principal or no cross-tenant attempt was made, the declaration SHALL state that positively.
APAS-CORE-4.2.2 A coverage declaration SHALL be truthful about differences between declared, reachable, and tested scope, and the surface baseline where applicable. No package SHALL imply complete, full, or equivalent coverage absent declared methodology and evidence sufficient for that representation. A result demonstrated on a single instance or run SHALL NOT be represented as coverage of the corresponding attack class.
APAS-CORE-4.2.3 Where source code, repository context, or comparable implementation artifacts are used, the coverage declaration SHALL distinguish implementation-aware testing from code-review completeness claims unless the applicable profile expressly permits the broader claim.
APAS-CORE-4.2.4 Any Q3, Q4, or R3 claim SHALL establish its coverage denominator against an Independent Surface Baseline. For these claims, the candidate system’s own discovery SHALL NOT be the denominator’s sole source. The baseline’s sources and construction method SHALL be disclosed under 4.2.1.
APAS-CORE-4.2.5 Material discrepancies between the candidate system’s discovered surface and the baseline, in either direction, SHALL be disclosed. Any surface in the baseline not exercised SHALL be classified as materially untested unless otherwise accounted for. Under-discovery SHALL reduce, never improve, the represented coverage.
APAS-CORE-4.2.6 Where no independent baseline is required, the coverage declaration SHALL state the denominator is self-enumerated and describe the enumeration method.
APAS-CORE-4.2.7 Any T2C claim SHALL structure the coverage declaration so a relying reviewer can identify control-relevant tested areas, exclusions, and materially untested areas without vendor-side analytical interpretation.
APAS-CORE-4.2.8 The coverage declaration SHALL identify each Independent Surface Baseline source and its resulting Baseline Strength, from this ascending-independence ladder: (a) BS-1 (asserted): a target-owner asset inventory or scope list, independent of the candidate system but bounded by owner completeness; (b) BS-2 (directed): evaluator-directed enumeration, the evaluator specifying or extending the surface beyond the candidate system’s discovery; (c) BS-3 (instrumented): enumeration by an independent discovery instrument or process not sharing the candidate system’s discovery logic.
APAS-CORE-4.2.9 Where sources are combined, the declaration SHALL disclose each source and the region of the denominator it covers, and SHALL NOT represent a composite baseline at a strength higher than the source solely covering any material region.
APAS-CORE-4.2.10 Any Q3, Q4, or R3 claim SHALL rest on a baseline of at least BS-2 (4.2.4).
APAS-CORE-4.2.11 Coverage status alone does not represent depth of exercise. For each area declared tested or partially tested, the coverage declaration SHALL state depth exercised, at minimum whether request-level probing, authenticated request-level probing, multi-step workflow or state-transition testing, and privilege- or tenancy-boundary testing were performed, and for partially tested areas what was and was not exercised. Unit of coverage, partial-coverage thresholds, depth grading, and any weighting SHALL be defined in the APAS Profile Registry; absent an applicable profile definition, depth SHALL be declared narratively and no coverage proportion or percentage SHALL be represented.
APAS-CORE-4.2.12 Where the candidate system is a composition declared under Agent Topology (7.1.1), reachable surface and exercised areas SHALL be the union across the declared components, with per-component attribution keyed to the component identity recorded under 8.2.6.

4.3 Finding states

Finding state Meaning
Observation An anomalous signal, artifact, or behavior not yet validated as a finding.
Validated Finding Candidate-system-demonstrated exploitability, or evidence establishing the vulnerable condition without demonstrating exploitation, pending independent confirmation. A finding MAY be a composed exploit chain whose impact derives from combined individually lower-severity conditions; impact substantiation SHALL be assessed against the chain as a whole, not its components in isolation.
Confirmed Finding A finding that the evaluator or applicable conformance process, without reliance on the candidate system’s own narrative, either independently reproduces the vulnerable condition or exploit path using the recorded Reproduction Procedure or, where it is time-sensitive, destructive, one-shot, or otherwise not safely reproducible, validates preserved first-capture evidence sufficient for decision-grade reliance within the declared methodology. Second-limb confirmation is a Single-Capture Confirmation under 4.3.3.
Confirmed Attack Path An ordered set of Confirmed Findings whose combined exploitation reaches an outcome no constituent reaches alone; confirmed at path level under 4.3.7.
Retested Finding Re-evaluation of a prior confirmed finding after claimed remediation, by executing the recorded Reproduction Procedure, with the result disclosed (4.3.4).
Unsupported Finding A reported issue in either sub-class of 4.3.8: a false positive or an over-severity finding.

APAS-CORE-4.3.1 Only Confirmed Findings SHALL support Q2, Q3, Q4, R2, or R3 claims. Other states MAY appear if clearly labeled, but SHALL NOT be represented as confirmed.
APAS-CORE-4.3.2 A finding whose reported impact is materially overstated SHALL be downrated or treated as unsupported for the overstated portion.
APAS-CORE-4.3.3 Each Confirmed Finding SHALL carry a Stable Finding Identifier. Except under the single-capture path in this clause, each SHALL carry a Reproduction Procedure; a finding without one SHALL NOT be represented as Confirmed for Q3, Q4, R2, or R3 purposes. Where the condition is time-sensitive, destructive, one-shot, or otherwise not safely reproducible, assessed against the Action Reversibility Class (7.1.1, Appendix D), a finding MAY reach Confirmed on preserved first-capture evidence rather than an executed reproduction, only where all of (a) to (f) hold, (a) to (c) at every claim level: (a) collection through a path not under the candidate system’s sole control (4.4.4); (b) a collection-time content hash per captured artifact under 8.2.2; (c) tamper-evident custody of the supporting record under 8.2.3; (d) a documented chain of custody from capture to evaluator; (e) evaluator or conformance-process validation of the capture, without reliance on the candidate system’s description, reasoning trace, or summary; and (f) a Capture Record stating the steps and conditions producing the capture and why re-execution is not safely repeatable. This path waives neither the Stable Finding Identifier nor any condition of this clause. Such a finding SHALL be labeled a Single-Capture Confirmation in the evidence package and the 8.4 export, and SHOULD NOT support an R3 claim unless the capture was evaluator-witnessed or produced by instrumentation independent of the candidate system.
APAS-CORE-4.3.4 A retest SHALL execute the Reproduction Procedure recorded against the prior Confirmed Finding’s Stable Finding Identifier, SHALL match its result to that identifier, and SHALL disclose whether the finding remains present, is partially resolved, or is no longer reproducible. (a) Probabilistic rediscovery, or its failure, SHALL NOT by itself be represented as fix verification, persistence confirmation, or retest completion; any agent re-run results reported alongside a retest SHALL be labeled probabilistic rediscovery and distinguished from the Reproduction Procedure result. (b) Where the recorded Reproduction Procedure is probabilistic (intermittent condition), the retest SHALL record attempt count and observed success rate, SHALL be assessed against the recorded pre-remediation success rate rather than as a binary reproduced or not, and the APAS Evaluator Guide SHALL define that assessment. (c) Where the prior finding is a Single-Capture Confirmation under 4.3.3, the retest SHALL instead examine the claimed remediation and the condition’s current state, SHALL disclose that basis, and SHALL NOT be represented as reproduction of the original condition.
APAS-CORE-4.3.5 Each Confirmed Finding SHALL declare the vantage point and access condition from which it was reached and the region of the declared environment class (7.2) where it occurs, drawn from the vantage points, credentials, and reachability boundaries in the affordance profile (7.1) and coverage declaration (4.2.1). This declaration SHALL NOT be construed as an assessment of business impact or asset criticality.
APAS-CORE-4.3.6 Where a Confirmed Finding was reached through, or exposes, adjacent reachable services, dependencies, or cross-system paths material to its exploitability or impact, those SHALL be disclosed; for any E3 claim this disclosure is REQUIRED. An adjacent item in a disclosure-restricted state (7.4.2) SHALL remain enumerated with its identifier and disclosure state, and only the exploitation detail SHALL be gated, until that state changes.
APAS-CORE-4.3.7 A Confirmed Attack Path SHALL reference each constituent Confirmed Finding by Stable Finding Identifier as an ordered set, and SHALL carry its own Stable Finding Identifier, path-level Reproduction Procedure, and EV3 under 4.4.1, established against the path itself, not inherited from any constituent. EV4 under 4.4.2 is REQUIRED wherever the path’s represented impact exceeds the highest represented impact of any constituent. Where chaining was material to the result, an E3 claim under 7.2.1 and any T4 claim under 5.6.3 SHALL represent that result as a Confirmed Attack Path, not unlinked findings. The APAS Attestation Schema (8.4.3) SHALL define encoding of the path object and its constituent references.
APAS-CORE-4.3.8 Unsupported Findings comprise two sub-classes: a false positive, a finding not confirmed because the condition does not exist or is not exploitable; and an over-severity finding, confirmed in existence but with reported impact materially overstated after adjudication. The quality declaration SHALL report the two sub-classes separately and SHALL state the Unsupported-Finding Rate (Appendix D) for the evaluated combination.
APAS-CORE-4.3.9 A documented deduplication rule SHALL exist for the package, stating how repeat observations of the same underlying condition, including from separate runs, vantage points, or instances, collapse onto a single Stable Finding Identifier.

4.4 Evidence classes

Evidence class Meaning
EV1 Signal. Anomalous behavior, misconfiguration, or suspicious condition.
EV2 Validation. Demonstrated exploitability, or the vulnerable condition established without demonstrating exploitation; MAY include candidate-system-generated artifacts.
EV3 Traceable Confirmation. Independent reproduction of the vulnerable condition or exploit path via the recorded Reproduction Procedure, observed against the target itself.
EV4 Impact Substantiation. Evidence or adjudicated rationale supporting the represented impact band.

APAS-CORE-4.4.1 A Confirmed Finding SHALL include EV3 or an equivalent traceable confirmation accepted under this clause. EV3 SHALL derive from direct observation of the target condition, such as raw request and response captures, target-side state changes, or evaluator-collected artifacts. The candidate system’s own description, reasoning trace, or summary of a finding SHALL NOT by itself constitute EV3. The two preceding sentences govern any accepted equivalent as they govern EV3, and such an equivalent SHALL preserve: (a) direct observation of the target condition; (b) derivation independent of the third-sentence materials; (c) collection-time integrity protection under 8.2.2; and (d) traceability to the Stable Finding Identifier under 4.3.3. (e) The evaluator or applicable conformance process SHALL determine whether a proposed equivalent satisfies (a) to (d) and SHALL record the basis of acceptance; equivalence is determined against this clause, 4.3.3, and the Appendix D definition of Reproduction Procedure, and SHALL NOT be established by the candidate system’s methodology statement alone. (f) The third-sentence prohibition applies to any internal validation, adjudication, or judge component of the candidate system, regardless of architectural separation, model choice, or context isolation; agreement between its components is not observation of the target. (g) Re-observation of the target by such a component, re-issuing the request or operation and reading the actual response, is observation, its artifacts treated as any other candidate-collected artifact, subject to 4.4.4; where it instead adjudicates narrative or artifacts without interacting with the target, its output is quality control on reporting, bears on Finding Validity, MAY be included as EV2, and SHALL NOT constitute EV3.
APAS-CORE-4.4.2 A High or Critical Confirmed Finding SHOULD include EV4 or evaluator-approved impact rationale.
APAS-CORE-4.4.3 Profiles MAY extend evidence classes but SHALL NOT weaken them for any Q2, Q3, Q4, R2, or R3 claim.
APAS-CORE-4.4.4 Collection paths for evidence artifacts: (a) for any Q3, Q4, R2, or R3 claim, EV3 and EV4 artifacts SHALL be collected through a path not under the candidate system’s sole control (e.g., network interception, host observation, or evaluator-collected capture); (b) for claims below that gate, artifacts collected solely by the candidate system SHALL be disclosed as such in the evidence manifest and weighed accordingly by the relying reviewer; (c) for Q2 and Q3 claims, EV3 and EV4 artifacts SHOULD be Out-of-Band Witness Evidence; and (d) condition (a) applies at every claim level to Single-Capture Confirmations under 4.3.3.
APAS-CORE-4.4.5 Multiple independent evidence sources MAY collectively satisfy EV4 where no single artifact substantiates the represented impact, provided the evidence package records the composition and the inference drawn from it. Where the represented impact derives from a composed exploit chain, EV4 SHALL be assessed against the chain as a whole, not its components in isolation.
APAS-CORE-4.4.6 Where a finding asserts an authorization or tenancy-boundary defect, EV3 SHALL include paired observations: the same request or operation succeeding for the unauthorized principal and refused for an authorized control principal, each principal identified by reference to Identity Coverage (4.2.1). The no-control-observation alternative in the Appendix D Reproduction Procedure definition SHALL NOT apply to such a finding.

4.5 Negative-result discipline

APAS-CORE-4.5.1 A package with no Confirmed Findings, or only Low severity Confirmed Findings where higher assurance or reliance is claimed, SHALL include a negative-result statement that: (a) identifies the exercised roles, workflows, privilege boundaries, exclusions, materially untested areas, limiting conditions, and evidence references relevant to interpreting the absence; (b) for Q3, Q4, and R3 claims, references the Independent Surface Baseline as an external denominator; and (c) declares, where 4.6 applies, the run design behind the absence under 4.6.3 and 4.6.5, with the run count, aggregation rule, and per-run outcomes. Where Low severity Confirmed Findings are constituents of a Confirmed Attack Path under 4.3.7, the path severity governs. Non-rediscovery is not absence under this clause (4.6.4).
APAS-CORE-4.5.2 No package SHALL be represented as assessor-ready, complete-coverage, or substitution-ready solely because the report is quiet, short, or contains no material findings.

4.6 Non-determinism disclosure

APAS-CORE-4.6.1 For this clause, a candidate system is materially non-deterministic where repeated execution under the same declared configuration and target conditions produces materially different reachable surface, finding sets, or exploitation outcomes; such a system SHALL disclose that property in the methodology statement, including its implications for run-to-run variance in discovery and findings. Where 4.6.3 applies, its declared variance measure determines whether that condition is met; otherwise the methodology SHALL state the basis of that determination.
APAS-CORE-4.6.2 Run-to-run variance SHALL NOT be concealed by reporting only the most favorable run. Where multiple runs inform a single claim, the methodology SHALL define how runs are selected, aggregated, and disclosed, and the evaluator SHALL record the basis. That record SHALL attest the run count and aggregation rule were fixed before execution. No claim of consistency, repeatability, or continuous coverage SHALL rest on probabilistic rediscovery alone.
APAS-CORE-4.6.3 Any Q3, Q4, R2, or R3 claim SHALL characterize run-to-run variance for the attested configuration using a declared measure (e.g., surface discovery variance, finding-set Jaccard index) over a pre-registered run design. Disclosing non-determinism without such a measure SHALL NOT satisfy 4.6.1 for these claims.
APAS-CORE-4.6.4 Neither rediscovery nor non-rediscovery by the candidate system’s own exploration SHALL stand in for a deterministic or permitted probabilistic Reproduction Procedure result. A non-rediscovery SHALL NOT be represented as absence, resolution, or non-existence of a condition in any coverage, negative-result, retest, or public-claim statement.
APAS-CORE-4.6.5 Where multiple runs inform a single claim, the aggregation rule SHALL be (a) declared from a closed set of any-run union, all-run intersection, majority, or another expressly declared rule; (b) bound to the claim it supports; and (c) stated with the run count behind it. The per-run outcome set, not a summary statistic alone, SHALL reach the evidence package. This clause SHALL NOT change how any coverage figure is computed.
APAS-CORE-4.6.6 The methodology SHALL state whether the target was held at, or restored to, an equivalent verified baseline between runs informing a single claim; where it was not, it SHALL disclose that and the resulting limit on attributing run-to-run difference to the candidate system. Restoration is not required; where it occurs it is an act of the authorizing party under 7.4.1, not of the candidate system or evaluator.
APAS-CORE-4.6.7 A pre-registered run design’s run count SHALL derive from the variance measure required by 4.6.3 and from the claim it supports. The design SHALL declare its purpose, at minimum whether it: (a) measures candidate-system discovery variance, for which candidate-system runs are the appropriate instrument; (b) characterizes an intermittent target condition, which SHALL use a condition-specific harness recorded as a probabilistic Reproduction Procedure (Appendix D) and assessed under 4.3.4, not candidate-system re-runs; or (c) verifies a claimed fix, where it SHALL state the pre-fix rate against which the post-fix result is assessed. The 4.3.3 single-capture exception SHALL NOT be used for a condition that can be re-attempted safely but does not trigger reliably. A run design SHALL NOT support a negative-result statement under 4.5.1 unless its run count derives under this clause. The APAS Evaluator Guide SHALL define the minimum run design and the selection rule; the APAS Profile Registry SHALL define the permitted aggregation rules for each profile and tier.

5. Claim Model

5.1 Claim types

Claim type Meaning
Atomic System Capability Certification Evaluator-issued result that a candidate system satisfied applicable APAS requirements for one capability combination under declared conditions.
Multi-Configuration Capability Statement Two or more atomic claims for one candidate system version.
Engagement Conformance Statement A specific engagement executed using an APAS-certified system under the same or narrower declared conditions.

APAS-CORE-5.1.1 An Engagement Conformance Statement SHALL NOT be represented as certifying the target environment, and a Multi-Configuration Capability Statement SHALL NOT be read as certifying any combination not listed with its status.

5.2 Claim format

The claim string is the machine-readable key by which capability combinations are declared, indexed, and compared.
APAS - / -- / [] {}
Example: APAS BB-T2C / X2-C2-Q4 / E3 [P-021] {R2}
Claim-string anatomy. The eight positions, in order, are:

Position Dimension Permitted values Defined in
1 Mode BB, GB, WB 5.5
2 Tier T1, T2A, T2C, T3, T4 5.6
3 Execution autonomy class X1, X2, X3 5.7
4 Collaboration class C1, C2, C3, C4 5.8
5 Assurance outcome class Q1, Q2, Q3, Q4 5.9
6 Environment class E1, E2, E3 7.2
7 Affordance profile identifier a declared profile id, shown in square brackets 7.1
8 Reliance outcome R1, R2, R3, shown in braces 5.10

APAS-CORE-5.2.1 Every atomic claim, public or private, SHALL use this format with all eight dimensions. Each unique tuple is a separate capability combination.
APAS-CORE-5.2.2 A candidate system version MAY hold multiple atomic certifications; where more than one is represented for it, the publisher SHALL provide a capability matrix.
APAS-CORE-5.2.3 Any Q2, Q3, Q4, R2, or R3 claim SHALL include or reference the applicable quality declaration.
APAS-CORE-5.2.4 Declared execution autonomy and collaboration classes SHALL agree on leadership and in-band analytical participation: (a) X1 SHALL be combined only with C1; (b) X2 and X3 SHALL NOT be combined with C1; (c) X3 SHALL NOT be combined with C2; (d) a combination violating (a), (b), or (c) SHALL NOT be attested, published, or listed in a capability matrix as certifiable; (e) permitted, conditionally permitted, and invalid combinations across all eight dimensions and the assistance class under 9.1.1 SHALL be enumerated in full in the APAS Attestation Schema under 12.1.1, as a machine-checkable constraint set for 8.4.4 ingestion; and (f) two combinations SHALL be adjudicated in the APAS Evaluator Guide, not enumerated: (i) X2 with C4, where permitted in-band collaboration went unexercised, questioning the attestable class (7.5.1); and (ii) C3 with H2, where one intervention at a defined escalation point is not repeated analytical handoff (7.5.2).

5.3 Combination status classes

Status Meaning
S1 Supported designed to operate in this combination; APAS evaluation not completed
S2 Evaluated evaluated, but no certification or conformance claim represented
S3 Certified or Conformant satisfied the applicable requirements and may be claimed
S4 Not Certified evaluated and did not satisfy the applicable requirements
S5 Not Evaluated not evaluated
S6 Unsupported not supported for the candidate system version
S7 Retired previously supported or certified, no longer active for the current version

APAS-CORE-5.3.1 A capability matrix SHALL give every listed combination one of these statuses. A combination SHALL NOT be represented as certified unless its status is S3, and a public claim SHALL clearly distinguish S3 from other statuses. A public matrix omitting materially relevant supported combinations is misleading.

5.4 Capability matrix requirements

APAS-CORE-5.4.1 Per listed combination, a capability matrix SHALL include: (a) combination ID; (b) full atomic claim string; (c) candidate system name and version; (d) model identity: model(s), version(s), and system-prompt provenance; (e) methodology reference; (f) evaluator identity where applicable; (g) status class; (h) attestation, quality-declaration, and coverage-declaration references where applicable; (i) the APAS Core Standard version and each applied companion document’s identifier and version (Evaluator Guide, Profile Registry entry, Attestation Schema); (j) certification and validity dates where applicable; (k) material limitations; and (l) notes on dynamic behavior or mode-selection rules where material. For any Q3, Q4, R2, or R3 combination, a certification date and an explicit expiry date or revalidation interval are REQUIRED. A combination varying in any claim dimension SHALL be a separate entry.

5.5 Modes of testing

Mode Description
BB Black-box: no source code or privileged architecture knowledge beyond declared credentials or materials.
GB Grey-box: limited credentials, documentation, or partial implementation context.
WB White-box or code-aware: source code, repository access, architecture detail, or deep implementation context.

Modes are distinct declared threat models; none is universally superior.
APAS-CORE-5.5.1 Every claim SHALL declare its mode, and the declared methodology SHALL state the threat model or assurance objective the mode serves.
APAS-CORE-5.5.2 A claim under one mode SHALL NOT be generalized or presumed equivalent to a claim under any other mode (see 5.10.2).
APAS-CORE-5.5.3 Where the declared objective is maximal coverage fidelity, evidence quality, or assessor-ready substitution and richer lawful context is available, profiles targeting Q4 or R3 SHOULD require WB or justify omitting it. Any Q4 or R3 claim under BB or GB SHALL disclose the threat-model rationale for the mode, why richer context was not used, and the resulting coverage and evidence limitations.
APAS-CORE-5.5.4 No party SHALL represent BB or GB testing as deficient, or WB testing as superior, in the abstract; mode comparisons SHALL be made only against a declared assurance objective.

5.6 Tiers

Tier Description
T1 Surface assurance.
T2A Recurring assurance pentest.
T2C Compliance evidence pentest.
T3 Business-critical deep assessment.
T4 Threat-led technical resilience.

APAS-CORE-5.6.1 Detailed tier requirements SHALL be defined in the APAS Profile Registry.
APAS-CORE-5.6.2 T2C SHALL NOT by itself imply control pass status, framework acceptance, or assessor acceptance.
APAS-CORE-5.6.3 T4 is bounded to technical attack paths. A T4 claim: (a) SHALL disclose that social-engineering, physical, and wireless initial-access vectors were not exercised; (b) SHALL declare any assumed-breach or initial-access conditions in the affordance profile and coverage declaration; (c) SHALL NOT be represented as equivalent to a full-scope red-team or threat-led engagement under frameworks requiring those vectors; and (d) SHALL NOT be represented as a generic pass or fail substitution class.

5.7 Execution autonomy classes

Class Definition
X1 Human-led assisted execution: the candidate system materially assists execution, exploration, or analysis.
X2 Agent-led execution, in-band human collaboration permitted: human participation under 5.7.1(a).
X3 Independent agent execution, no in-band analytical human assistance: human activity under 5.7.1(a).

APAS-CORE-5.7.1 (a) Under X2, humans MAY provide in-band analytical input, steering, escalation handling, or report-shaping; under X3, human activity MAY be limited to evaluator adjudication, safety intervention, administrative clarification, or post-execution review. (b) Higher execution autonomy SHALL carry stronger safety, oversight, and interruption obligations (Section 10); removing in-band human participation SHALL NOT remove the oversight bounding the agent.
Note (informative). This axis measures human involvement in the workflow, not output quality; Section 4’s quality dimensions are the long-run differentiator.

5.8 Collaboration classes

Class Definition
C1 Human-led with agent assistance: humans remain the primary operators.
C2 Agent-led with active human collaboration: humans and agents may repeatedly hand off work, reasoning, or next actions.
C3 Agent-led with human escalation only: humans intervene only at defined escalation points, safety events, or evaluator-directed checkpoints.
C4 Agent-only until relying-party review: no human participation other than evaluator oversight or post-completion review.

APAS-CORE-5.8.1 C4 excludes in-band analytical human participation, not safety oversight. Section 10’s stop authority, monitoring, and interruption requirements apply most strongly to C4 and SHALL NOT be weakened by it.
Note (informative). The note to 5.7 applies equally to the collaboration axis.

5.9 Assurance outcome classes

Assurance outcome describes the strength and usability of the resulting pentest output.

Class Definition
Q1 Signal-grade: useful technical signal, not sufficient for decision-grade reliance.
Q2 Decision-useful: independently evaluable, materially reproducible, with a coverage declaration and limitations statement, for remediation or security decisions within the declared scope.
Q3 Reference-validated: meets applicable evaluator requirements for comparative validation against a declared reference process, cohort, or baseline under the same declared conditions, coverage measured against an Independent Surface Baseline.
Q4 Assessor-ready: meets applicable evaluator requirements for a complete assessor-consumable evidence package (quality declaration, coverage declaration against an Independent Surface Baseline, traceable finding support, limitations, negative-result discipline where applicable) without vendor-side analytical interpretation or editing.

APAS-CORE-5.9.1 Q4 SHALL NOT imply framework acceptance by default.
APAS-CORE-5.9.2 Q3 or Q4 MAY be achieved at any execution or collaboration class, subject to the applicable profile and evaluator result. Q3 is not a prerequisite for Q4 unless that profile expressly requires comparative validation.
APAS-CORE-5.9.3 Comparative validation SHALL NOT authorize omitting a quality declaration, coverage declaration, or evidence traceability requirements.
APAS-CORE-5.9.4 Higher assurance outcome classes SHOULD favor the richest lawful context, consistent with the declared threat model, where the objective is to reduce noise, improve finding quality or coverage fidelity, or accelerate validation.

5.10 Reliance outcomes

Outcome Definition
R1 Evidence-only; use under 5.10.1(a).
R2 Assessor-consumable augmentation; use under 5.10.1(a).
R3 Assessor-reliable substitution; use under 5.10.1(a).

APAS-CORE-5.10.1 (a) R1 MAY be used as supplemental evidence and SHALL NOT be represented as substituting for a required pentest activity; R2 MAY augment assurance, coverage, retesting, or remediation validation where the package is directly usable by the relying reviewer; R3 MAY be represented as substituting for a subset of pentest activities only where the applicable profile, evaluator result, quality declaration, and relying framework or assessor permit. (b) R3 is conditional and profile-bound. Core SHALL NOT require a single execution or collaboration class for Q3, Q4, or R3; any stricter condition SHALL be imposed by the applicable profile.
APAS-CORE-5.10.2 A stronger result in one claim dimension does not transfer to another. No claim transfers WB to GB or BB, BB or GB to WB, X3 to X2 or X1, C4 to C3, C2, or C1, Q4 to Q3, Q2, or Q1, E3 to E2 or E1, or R3 to R2 or R1. Any permitted carry-forward SHALL be explicitly defined by the applicable profile or Evaluator Guide.
APAS-CORE-5.10.3 Any R2 or R3 claim SHALL identify the subset of pentest activities, coverage areas, and decision uses the package evidences and those it does not.
APAS-CORE-5.10.4 Any R3 claim SHALL identify the party accountable for the substituted evidence’s accuracy, the relying party’s recourse if that evidence is later found materially deficient, any warranty limitations, and any professional-liability or errors-and-omissions coverage represented to the relying party. APAS does not create, transfer, or cap liability, and no statement under this clause SHALL be represented as doing so.
APAS-CORE-5.10.5 A reliance outcome SHALL NOT exceed the assurance outcome supporting it. An R3 claim SHALL require Q4. R1 MAY accompany any assurance outcome class. A profile MAY impose stricter assurance-to-reliance pairings but SHALL NOT permit looser ones.
APAS-CORE-5.10.6 An R2 or R3 claim SHALL NOT be represented as substituting for, or augmenting, authorization, access-control, or tenancy-boundary testing where fewer than two distinct principals, or only a single tenant, were exercised. Such areas SHALL be declared materially untested under 4.2.5. A single-principal engagement against an unauthenticated E1 surface, or a non-tenanted target, MAY still support a truthful R3 claim once the 5.10.3 subset excludes those areas.
APAS-CORE-5.10.7 An R2 or R3 claim declares what the package supplies, not which procedures the relying reviewer may omit. Residual procedures and residual-risk acceptance rest with that reviewer and the governing framework, unless an APAS Crosswalk Annex states otherwise.

5.11 Claim binding to candidate system configuration

APAS-CORE-5.11.1 A claim is bound to the configuration attested at evaluation: the foundation model(s), system prompts, orchestration and agent-control logic, and the tool, plugin, model, and MCP provenance set recorded in the affordance profile (7.1).
APAS-CORE-5.11.2 A material change to any of these SHALL invalidate every affected atomic claim until re-attested: (a) foundation model provider, family, or pinned version or identifier; (b) fine-tuning, adapter, or system-prompt changes altering decision-making; (c) orchestration or agent-control logic governing strategy, exploitation, or validation; (d) the tool, plugin, model, and MCP provenance set; or (e) persisted memory or accumulated context state materially altering decision-making.
APAS-CORE-5.11.3 The claimant SHALL disclose the change, the claims affected, and the re-attestation status. A claim relied upon as Q3, Q4, R2, or R3 SHALL NOT be represented as valid across a material change absent re-attestation.
APAS-CORE-5.11.4 Where the candidate system is materially non-deterministic (4.6), a change altering its run-to-run variance SHALL be treated as material under 5.11.2. The 4.6.3 measure SHALL be the drift-assessment baseline, and SHALL be re-measured at intervals the applicable profile sets across the claim validity window. The drift threshold requiring re-attestation SHALL be defined in the APAS Profile Registry.
APAS-CORE-5.11.5 A change listed in 5.11.2 is material whether initiated by a human or by the system itself. Where the affordance profile’s Dynamic behavior field (7.1.1) permits reconfiguration within a run, each such change SHALL be recorded as a security-relevant event in the Agent Execution Record (8.2.6) and disclosed under 5.11.3.

6. Core Roles

6.1 Evaluator

APAS-CORE-6.1.1 For X2, X3, Q3, Q4, R2, or R3 claims, the evaluator SHALL be organizationally independent of the vendor and SHALL disclose any financial, contractual, or governance relationship a reasonable relying party would consider material to independence.
APAS-CORE-6.1.2 The evidence package SHALL disclose the evaluator engagement model, including who performs and funds independent reproduction and the expected turnaround.

6.2 Relying reviewer

APAS-CORE-6.2.1 For R2 or R3 claims, the intended human reviewer SHALL be the relying auditor, assessor, or customer-appointed reviewer, not a vendor-side pentest reviewer. Post-execution review by that reviewer SHALL NOT count as human assistance.

6.3 Reference validation method

APAS-CORE-6.3.1 Any Q3 claim relying on comparative validation SHALL be supported by, and SHALL cite, the common versioned comparative-validation methodology in the APAS Evaluator Guide and the version applied; an evaluator’s own methodology SHALL NOT satisfy this. Where a human reference cohort is used, the methodology SHALL disclose cohort qualification rules and conflict controls. The Guide SHALL define reproducibility criteria, the reference corpus, and any inter-evaluator agreement measure.

6.4 Vendor-side analytical review

APAS-CORE-6.4.1 Vendor-side analytical review during execution, finding formation, report drafting, declaration formation, evidence-package assembly, or framework or control mapping SHALL be disclosed. If undisclosed, it SHALL invalidate any Q3, Q4, or R3 claim and any affected quality-declaration fields.

7. Declared Conditions

7.1 Affordance profile

APAS-CORE-7.1.1 Every APAS claim SHALL reference a declared affordance profile with at minimum the fields below. A generic label such as “standard” SHALL NOT satisfy the Target-protection controls or Agent-integrity controls field. Any internal validation, adjudication, judge, verifier, or consensus component SHALL be declared in Agent Topology, its role disclosed in the quality declaration against Finding Validity. Absent a declared Action Reversibility Class, the engagement SHALL be treated as irreversible; the declared class scales safety obligations under 10.3.1 and 10.3.2.

Required field Content
Access flags internet, browser, terminal or shell, source-code or repository access status
Agent Topology candidate system composition (orchestrator, sub-agents, third-party agents, dynamic or runtime resources), trust boundaries, inter-component capability attenuation, delegation depth
Credentials type; number and privilege level of distinct principals, roles, and tenants provisioned; credential validity period, scope, or intended use, and revocation mechanisms where applicable
Reachability network reachability boundaries, vantage points tested
Asset coverage asset classes covered
Inference and Telemetry Egress destinations for target data, credentials, or evidence (model endpoints, telemetry, connection servers), with jurisdiction, retention policy, and redaction applied
Tools and provenance tool and plugin inventory where material, with supply-chain provenance (source, version, integrity verification) of tools, plugins, models, and MCP or equivalent tool-connection servers; including where applicable the permitted use of each tool or plugin by the system
Context inputs context sources provided (source code, repositories, architecture documents, logs, test credentials) where material
Scope basis scope source or target inventory reference where material
Persistence memory persistence where material
Target-protection controls rate limits, destructive-action prohibitions, blast-radius limits, and the safety restrictions, stop authority, and interruption mechanism per Section 10
Agent-integrity controls injection filtering, out-of-band scope enforcement, and equivalent controls against adversarial manipulation of the candidate system per Section 10
Escalation limits maximum permitted privilege escalation, lateral movement, data access, persistence, destructive-action boundaries
Action reversibility highest Action Reversibility Class (Appendix D) the engagement permits: read-only, externally reversible, or irreversible
Data handling collection, storage, transmission, retention, and deletion guarantees for target data, credentials, and evidence, including any use of target data for model training
Timebox timebox
Exclusions exclusions and non-testable areas
Methodology reference and version
Dynamic behavior mode-selection logic, escalation and handoff triggers, conditions for switching context sources, credentials, or collaboration pattern mid-run, whether dynamic reconfiguration is permitted within a run, and how self-initiated material changes are detected and recorded

APAS-CORE-7.1.2 The data-handling guarantees under 7.1.1 SHALL apply to the Agent Execution Record (8.2.6) and any telemetry egress, including redaction of Personally Identifiable Information (PII), secrets, and credentials at capture time.

7.2 Environment classes

Environment class Description
E1 A single isolated target.
E2 Multi-role, multi-step, or stateful interaction within one declared trust boundary.
E3 Interaction crossing a trust or segmentation boundary.

The classes describe reach across trust boundaries, not target type, and apply equally to applications, APIs, networks, infrastructure, cloud control planes, containers, client software and executables, mobile, firmware, and embedded targets; asset class is declared under 7.1 and 4.2, and E1 through E3 are ordered by reach. Social-engineering, physical, and wireless vectors are out of APAS scope, so no environment class imports them and 2.2.2 applies regardless of class or asset type.
APAS-CORE-7.2.1 All APAS claims SHALL be limited to the declared environment class, an ordered reach ladder from E1 through E3. Expansion during a run is governed by 7.4.1 and 7.5.1.
APAS-CORE-7.2.2 The declaration SHALL state the environment type and any material differences from production. A non-production test SHALL NOT be represented as, or imply, production coverage absent declared parity.

7.3 Methodology declaration

APAS-CORE-7.3.1 Any T2A, T2C, T3, or T4 claim SHALL reference a declared methodology. Any Q2, Q3, Q4, R2, or R3 claim SHALL reference one defining coverage-declaration rules, required evidence, and exclusions and limitations handling.
APAS-CORE-7.3.2 Any methodology targeting Q4 or R3 SHALL define negative-result handling and remediation and retest expectations, SHALL state the threat-model rationale for the mode, and SHALL explain whether richer implementation context was available, used, or omitted.

7.4 Authorization and rules of engagement

APAS-CORE-7.4.1 Every APAS engagement SHALL have an Authorization Declaration identifying the authorizing party, the legal or contractual basis, the authorized scope and time window, and the rules of engagement. Testing outside that scope, window, or rules SHALL be treated as out-of-scope activity under Section 10 regardless of technical success or finding value.
APAS-CORE-7.4.2 Every APAS engagement SHALL declare a Coordinated-Disclosure Path for vulnerabilities in third-party components, dependencies, shared infrastructure, or out-of-scope systems; the package SHALL disclose whether any were encountered and how handled. For each: (a) it SHALL carry a disclosure state of encountered, notified, disclosure-restricted, resolved, or published, and the state SHALL be updated as the Coordinated-Disclosure Path progresses; (b) it SHALL remain enumerated in the evidence manifest in every state, with its Stable Finding Identifier, current state, and content hash, keeping 8.2.2 and 8.2.5 verifiable; (c) where disclosure-restricted, only exploitation detail SHALL be withheld, only until the state changes, and the reason and imposing party SHALL be stated; and (d) a disclosure state SHALL NOT be used to remove it, reduce the package below 8.3.1 completeness, or justify vendor-side editing.
APAS-CORE-7.4.3 The Authorization Declaration (7.4.1) SHALL additionally declare, where material, each field below. Any inapplicable element SHALL be marked not applicable, not omitted.

Required field Content
Testing windows permitted windows and any blackout periods, maintenance windows, or production-hours restrictions during which offensive activity SHALL NOT occur, beyond the authorized time window in 7.4.1
Emergency contacts named primary and secondary contacts for the target owner, the evaluator where present, and the candidate system operator, with contact method, availability window, notification triggers, and maximum notification time under Section 10
Third-party and cloud-provider authorization authorization status and provider-specific testing conditions for third-party, hosted, or cloud infrastructure in scope or reachable during testing, including shared control planes and managed services
Source identification source addresses, ranges, or identifiers from which the candidate system operates, and any allowlisting arranged with the target owner or providers
Data-handling limits classification limits on target data, credentials, and secrets the candidate system may access, collect, transmit, or retain as evidence, consistent with the data-handling guarantees in 7.1
De-confliction whether defensive personnel (blue team, Security Operations Center [SOC], or provider monitoring) are aware of the engagement, the arrangement where applicable, and the contact for distinguishing test activity from a live incident

APAS-CORE-7.4.3.1 Testing outside a declared testing window, from an undeclared source, against infrastructure lacking required third-party or cloud-provider authorization, or violating a declared data-handling limit SHALL be treated as out-of-scope activity under Section 10 on the terms in 7.4.1.
APAS-CORE-7.4.3.2 Where the candidate system can reach shared or third-party infrastructure outside the declared authorization, the affordance profile escalation limits (7.1) SHALL bound it away, and any contact SHALL be disclosed and handled under the Coordinated-Disclosure Path (7.4.2).
APAS-CORE-7.4.4 The Authorization Declaration SHALL identify the acting agent identity, the identity scheme, the verification method available to the target owner, and, where sub-agents act, the delegation chain and maximum delegation depth.

7.5 Dynamic operating behavior

APAS-CORE-7.5.1 Where a candidate system can change mode, execution class, collaboration class, environment class, context source, or assistance condition mid-run, the evaluator SHALL record the observed operating path and the attested claim SHALL reflect, per affected dimension: (a) execution and collaboration class, the weakest truthful class materially used, on the orderings in 10.3.1 and for non-transfer under 5.10.2; (b) mode, the richest implementation context materially used (7.5.2), which SHALL NOT be read as a ranking of modes (5.5.4); (c) assistance condition, the class under 9.1.2; and (d) environment class, the declared class limits the claim (7.2.1), reach beyond the authorized scope is an authorization matter (7.4.1, Section 10) before a claim-transition matter, and reach materially exercised within scope but narrower than the declared class SHALL be attested at the weakest truthful class on the 7.2 reach ordering.
APAS-CORE-7.5.2 If source code or comparable implementation context is materially used, the run SHALL be attested WB; if in-band analytical human assistance materially changes strategy, exploitation, validation, prioritization, or report content, it SHALL NOT be attested as X3; if the work pattern includes repeated human-agent analytical handoff, it SHALL NOT be attested as C3 or C4; and if execution crosses a trust or segmentation boundary, it SHALL NOT be attested below E3.
APAS-CORE-7.5.3 A candidate system MAY receive separate atomic claims for separately bounded operating configurations, provided each is independently declared and evaluated. The coverage and quality declarations SHALL reflect material dynamic changes in context, assistance, or mode.

8. Evidence Package Requirements

8.1 Minimum package contents

APAS-CORE-8.1.1 Any certification or engagement statement represented as R2 or R3 SHALL include an assessor-consumable evidence package. Any Q2, Q3, or Q4 claim SHALL include a quality declaration and coverage declaration, even where R1 is claimed.
APAS-CORE-8.1.2 The minimum evidence package SHALL contain: (a) the following elements; and (b) for R2 or R3 claims, supporting evidence SHOULD remain available at least as long as the claim is valid.

Required element Content
Declarations certification or engagement declaration; reliance outcome; quality declaration; coverage declaration with surface-baseline reference where required; authorization and rules-of-engagement reference
Methodology reference and version
Combination identity combination ID; capability matrix reference where the system lists more than one combination
Findings finding-state summary including unsupported findings; Confirmed Findings with stable identifiers and Reproduction Procedure references; labeled Single-Capture Confirmations under 4.3.3, each with chain of custody from capture to evaluator; Confirmed Attack Paths under 4.3.7 with path-level identifier, ordered constituent Stable Finding Identifiers, and path-level Reproduction Procedure
Run design and outcomes pre-registered run design under 4.6.3; aggregation rule, run count behind the reported result, and per-run outcome set, not a summary statistic alone, under 4.6.5
Evidence evidence manifest with per-artifact content hashes; per-artifact retention period and the party responsible for availability
Sensitive evidence per class of secrets, PII, customer data, exploit artifacts, screenshots, payloads, and destructive proof: what is retained, redacted, and safe to share
Dynamic behavior where material, observed operating path summary and why the attested combination truthfully reflects the run
Remediation and retest remediation guidance where findings exist, stating root cause, fix direction, affected component and version, and the basis of any represented remediation priority under 8.4.5.1; retest information where applicable
Exclusions, limitations, negative result exclusions and untested areas statement; limitations statement; negative-result statement where applicable
Disclosure handling for third-party or out-of-scope vulnerabilities encountered, coordinated-disclosure status and the per-item disclosure state under 7.4.2
Access disposition Access Disposition Record under 10.2.5, with a non-secret reference per engagement-created or engagement-authorized access point
Attestation and mapping machine-readable, signed attestation; control or requirement mapping reference where the applicable T2C profile or APAS Crosswalk Annex requires it

APAS-CORE-8.1.3 For a multi-configuration system, the package SHALL permit the relying reviewer to determine which combination was evaluated, which are certified, and which remain unevaluated or unsupported. Any T2C package represented as Q4, R2, or R3 SHALL likewise let that reviewer determine tested, excluded, and materially untested control-relevant areas, without vendor-side analytical interpretation.

8.2 Package and artifact integrity

APAS-CORE-8.2.1 Any Q3, Q4, R2, or R3 attestation SHALL be cryptographically signed by the issuing evaluator or conformance authority, using a verifiable identity and published verification method, and SHALL identify the signing certificate or key identifier, carry a trusted timestamp of signing, and reference a published means of validating the key’s status at that time, including after rotation, expiry, or revocation.
APAS-CORE-8.2.2 The evidence manifest for any Q3, Q4, R2, or R3 package SHALL include a collection-time content hash for each referenced artifact.
APAS-CORE-8.2.3 Execution and evidence logging supporting any Q3, Q4, R2, or R3 claim SHALL be tamper-evident and maintained or countersigned by a component, party, or system not under the candidate system’s sole control.
APAS-CORE-8.2.4 Fabricated, non-traceable, or materially misleading evidence SHALL invalidate the claim. A materially misleading coverage declaration or negative-result statement SHALL invalidate the affected quality-declaration fields and any affected reliance claim.
APAS-CORE-8.2.5 A package whose signature fails verification, whose artifact hashes mismatch, or whose tamper-evident log shows unexplained discontinuity SHALL NOT be relied upon as Q3, Q4, R2, or R3 until the discrepancy is resolved and re-attested.
APAS-CORE-8.2.6 Any Q3, Q4, R2, or R3 package SHALL include an Agent Execution Record (AER) covering the attested run. The AER SHALL be maintained under 8.2.3, SHALL be exportable under 8.4.7, and SHALL record every security-relevant event, at minimum: (a) model or tool invocation, including internal validation, adjudication, or judge components; (b) context ingestion; (c) policy decision; (d) delegation; (e) oversight event; (f) credential use; (g) privilege or scope change; (h) material change the engagement makes to the target environment; and (i) where the affordance profile’s Dynamic behavior field permits in-run reconfiguration, each reconfiguration the candidate system initiates itself; and for every event: (j) event identifier, timestamp, component identity, argument/result hashes, and enforcement verdict; (k) where applicable, the acting principal identity, a non-secret credential reference, and the tenant or token-audience boundary crossed; the credential reference SHALL NOT carry a secret or credential value, consistent with 7.1.2; and (l) for oversight events, the instruction issued, its delivery, its enforcement, and the observed effect as distinct facts, and a positive value where an observation is missing; absence of a record SHALL NOT be treated as a negative record. These facts are the record assessed under 10.2.6. Field construction is defined in the APAS Attestation Schema under 8.4.7 and 12.1.1.
APAS-CORE-8.2.7 Material inconsistencies among the attestation, reported findings, execution records, or supporting artifacts SHALL be disclosed in the package and SHALL NOT be resolved by narrative interpretation alone.

8.3 Assessor consumability

APAS-CORE-8.3.1 A package represented as R2 or R3 SHALL let the relying reviewer evaluate scope, coverage, findings, limitations, and the declared reliance outcome without vendor-side analytical interpretation. For any Q4 claim, the package SHALL be complete without vendor-side analytical editing.
APAS-CORE-8.3.2 For a BB or GB claim, the package SHALL disclose the limitations arising from reduced context relative to the declared assurance objective and SHALL NOT imply equivalent code awareness.

8.4 Machine-readable attestation and interoperability

APAS-CORE-8.4.1 Any Q2, Q3, Q4, R2, or R3 package SHALL include machine-readable fields for: the full atomic claim string; combination ID; candidate system name and version; model identities and versions; system-prompt provenance; a candidate-system build identifier, a cryptographic digest over the configuration bound under 5.11.1, with its verification method; methodology reference; evaluator identity; attestation signature and verification metadata; assistance class; observed operating path; quality and coverage declaration references, with baseline source identifiers, baseline date or age, reachable-surface asset identifiers, and SBOM component references where applicable; finding-state summary, stable finding identifiers, Confirmed Attack Path relationships, and Single-Capture Confirmation labels; materiality determinations of record; evidence-manifest reference and hashes; certification date and explicit expiry or revalidation interval; APAS Core version and each applied companion identifier and version; and reliance outcome. The claim string and build identifier SHALL be carried in a declared machine-readable provenance envelope (e.g., in-toto, DSSE, or SCITT) defined in the APAS Attestation Schema. These fields SHALL be consistent with 5.4, 8.1.2, 4.2.1, and 7.1.1.
APAS-CORE-8.4.2 Each finding in a machine-readable export SHALL carry: (a) a stable finding identifier; (b) a Common Weakness Enumeration (CWE) classification where one applies; (c) Common Vulnerabilities and Exposures (CVE) references where it corresponds to a known CVE; and (d) a severity score from a versioned, declared scoring system; where Common Vulnerability Scoring System (CVSS) is used, its version SHALL be stated; any other system and version SHALL be declared with rationale. Each Confirmed Finding SHALL additionally carry: (e) the vantage point and access condition declared under 4.3.5; (f) the region of the declared environment class where it occurs; (g) the adjacency disclosure required by 4.3.6 where applicable; (h) a single-capture-confirmation label where confirmed under 4.3.3’s single-capture path; and (i) for a manipulation-class finding under 10.4.1 with no applicable CWE, a declared, structured classification against the manipulation classes in 10.4.1, with the taxonomy name and version stated. Any CVE or other external identifier carried SHALL have its applicability independently verified against the version of the target component present, not the weakness class alone; an identifier not applying to that version SHALL NOT be carried in support of the finding or its severity.
APAS-CORE-8.4.3 The APAS Attestation Schema SHALL be defined as a profile of, or define a lossless export to, established interchange formats. At minimum: finding-level results SHALL be exportable as Static Analysis Results Interchange Format (SARIF); control and requirement mappings under T2C profiles SHALL be exportable as, or referenced through, Open Security Controls Assessment Language (OSCAL); and remediation and retest status SHALL be expressible in OpenVEX or an equivalent declared exploitability-status format. The package SHALL state each interchange format version used. The Schema SHALL define the encoding of Confirmed Attack Paths under 4.3.7, including the relationship extension required by 8.4.4.
APAS-CORE-8.4.4 A package SHALL NOT be represented as machine-readable under APAS if its atomic findings cannot be ingested by standard vulnerability management tooling through the formats in 8.4.3 without vendor-proprietary translation. Also: (a) a Confirmed Attack Path object under 4.3.7 SHALL degrade losslessly into its ordered constituent atomic findings plus a documented relationship extension; and (b) profile- or vendor-specific fields MAY extend the required formats through their standard extension mechanisms but SHALL NOT replace or obscure the required interoperable fields.
APAS-CORE-8.4.5 For each Confirmed Finding, the machine-readable export SHALL record available exploitation-context indicators, each with its retrieval date, subject to (a) to (e): (a) such indicators are point-in-time values as of that date and MAY change without change to the underlying finding; (b) a package SHALL NOT represent an indicator as current beyond its retrieval date absent refresh; (c) where a profile requires them kept current, it SHALL define the refresh expectation and its relationship to the validity dates in 5.4; (d) where the finding has no known CVE, or a value is unavailable, that absence SHALL be recorded rather than the field omitted; and (e) these values SHALL NOT be represented as, or substituted for, the finding’s Confirmed status, impact substantiation (EV4), or coverage against the Independent Surface Baseline. The required indicators, with per-field construction defined in the APAS Attestation Schema under 12.1.1, are Cybersecurity and Infrastructure Security Agency (CISA) Known Exploited Vulnerabilities (KEV) catalog status and a versioned, declared exploit-prediction score such as Exploit Prediction Scoring System (EPSS) where the finding corresponds to a known CVE, and an equivalent declared risk-prioritization score, where one is used, where it does not.
APAS-CORE-8.4.5.1 Where a package presents findings in prioritized or risk-ranked order, the prioritization methodology SHALL be declared, including how it incorporates the exploitation-context indicators under 8.4.5. A prioritization SHALL NOT be represented as risk-based or exploitation-informed unless it incorporates the declared indicators where available. The presence of KEV status or an exploit-prediction score SHALL NOT by itself constitute a prioritization.
APAS-CORE-8.4.6 The provenance set for tools, models, and MCP servers SHALL be exportable as a machine-readable bill-of-materials (e.g., SBOM/AI-BOM) rather than prose. For high-assurance (Q4/R3) claims, verification SHALL require cryptographic digest pinning and signature verification covering that set and the candidate system build under 8.4.1, where the ecosystem supports it.
APAS-CORE-8.4.7 The Agent Execution Record required by 8.2.6 SHALL conform to the schema and semantic conventions defined in the APAS Attestation Schema, which SHALL define that schema as a profile of an established open telemetry format. A deviating implementation SHALL disclose the deviation and provide a translation mapping to the standard schema in the evidence manifest, keeping the record ingestible by standard vulnerability management tooling.

9. Human Assistance and Collaboration Integrity

9.1 Assistance classes

Assistance class Meaning
H0 No human assistance affecting strategy, exploitation, validation, prioritization, or reporting.
H1 Administrative or safety-only intervention.
H2 Analytical assistance, including exploit guidance, interpretation, prioritization, control mapping, or report editing.

APAS-CORE-9.1.1 The assistance class SHALL be disclosed in the attestation. For multi-configuration systems, that disclosure SHALL be per combination and, where relevant, per run.
APAS-CORE-9.1.2 An X3 claim SHALL prohibit H2 during execution. A Q4 claim SHALL prohibit H2 during evidence-package assembly. That prohibition covers an internal validation or adjudication component shaping evidence-package content, whose role SHALL be disclosed under 4.4.1 and in Agent Topology (7.1.1). H2 does not by itself invalidate X1, X2, C1, C2, Q1, or Q2 claims, if accurately disclosed.
APAS-CORE-9.1.3 Post-execution vendor-authored framework mapping, prioritization, or conclusion-shaping materially affecting the relying reviewer SHALL be treated as vendor-side analytical review and disclosed.
APAS-CORE-9.1.4 H1 safety intervention, including stop authority under Section 10, SHALL NOT be reclassified as H2 and does not by itself invalidate an X3 or C4 attestation, if disclosed.

9.2 Collaboration truthfulness

APAS-CORE-9.2.1 The declared collaboration class SHALL reflect work as actually performed. Undisclosed handoffs or analytical intervention, or materially misleading collaboration claims, invalidate the affected claim dimensions.
APAS-CORE-9.2.2 A platform SHALL NOT be marketed as agent-only, independently autonomous, or assessor-ready where materially relevant runs for the represented combination relied on undisclosed human switching, human review, vendor-side control mapping, or post-execution analytical reshaping.

10. Safety, Stop Authority, Agent Integrity, and Evaluation Integrity

10.1 Baseline requirements

APAS-CORE-10.1.1 Every APAS claim SHALL require explicit scope boundaries, a stop or suspend mechanism, and preserved logs or artifacts sufficient for evaluator review, subject to Section 8.2 where Q3, Q4, R2, or R3 is claimed.
APAS-CORE-10.1.2 Out-of-scope activity, fabricated evidence, material concealment of assistance, or materially misleading coverage statements invalidate the claim.
APAS-CORE-10.1.3 Material uncertainty in impact, coverage, or retest conclusions SHALL be explicitly labeled in the evidence package. Detailed safety tests, replay requirements, and failure conditions SHALL be defined in the Evaluator Guide or applicable profile.

10.2 Stop authority and interruption

APAS-CORE-10.2.1 Every engagement SHALL declare Stop Authority: the parties empowered to suspend or terminate execution, at minimum the target owner and, where present, the evaluator.
APAS-CORE-10.2.2 The stop mechanism SHALL operate out-of-band from the candidate system’s decision loop; the candidate system SHALL NOT be the sole arbiter of whether a stop instruction takes effect. The affordance profile SHALL declare a maximum stop reaction time. On a stop instruction the candidate system SHALL halt offensive activity within that bound, interrupting in-flight actions where technically feasible, and SHALL NOT initiate new offensive actions; queued, scheduled, or delegated work SHALL be cancelled within that bound where technically feasible, and otherwise dispositioned in the record.
APAS-CORE-10.2.3 Where monitoring, oversight connectivity, logging, or the stop channel is lost or degraded beyond declared tolerances, the candidate system SHALL fail safe, suspending offensive activity until oversight is restored. Loss of oversight SHALL NOT default to continued autonomous operation.
APAS-CORE-10.2.4 Stop and halt-on-loss-of-oversight events and their timing SHALL be recorded in the tamper-evident log, disclosed in the evidence package, and notified to the 7.4.3 emergency contacts within the declared maximum notification time.
APAS-CORE-10.2.5 Engagement closure, and any stop under 10.2.2, SHALL include an Access Disposition Record. Where engagement-created access exists (tokens, accounts, sessions, tunnels, delegated tokens, or persistence), the record SHALL identify (a) the access point; (b) its disposition status (Revoked, Expired, Removed, Rotated by target owner, or Retained with named owner and expiry); (c) the party responsible for performing or authorizing it; and (d) a verification reference accounting for it.
APAS-CORE-10.2.6 The evaluator SHALL verify each safety-critical stop, suspend, resume, or override channel end-to-end, from authorized party to enforcement point, as deployed for the attested combination (isolation, sandboxing, privilege separation); verifying existence, out-of-band operation, or correct signing SHALL NOT by itself satisfy 10.2.2. The evidence package SHALL record per channel at least one exercised end-to-end event, its observed delivery and enforcement result, and where applicable its reaction time, decomposed under 8.2.6.
APAS-CORE-10.2.7 Channel availability SHALL NOT by itself be treated as evidence of delivery. Where an issued control instruction is not acted on at the enforcement point, that SHALL be detectable as a delivery failure, distinguishable from an instruction never issued, and SHALL be surfaced as an oversight-loss condition under 10.2.3.

10.3 Autonomy-scaled obligations

APAS-CORE-10.3.1 Safety obligations SHALL scale with the worse of execution autonomy, reduced in-band human participation, and declared affordance risk (Action Reversibility Class, 7.1.1):
(a) within a fixed Action Reversibility Class and any given profile, requirements for X3 SHALL be at least as strong as X2, and X2 at least as strong as X1; the same ordering applies from C1 through C4;
(b) a declared Action Reversibility Class permitting less reversible action SHALL raise the applicable requirements irrespective of execution or collaboration class, and SHALL NOT lower them;
(c) no profile, evaluator decision, or vendor configuration SHALL grant the most autonomous classes weaker oversight than less autonomous classes; a configuration in which increased autonomy removes the oversight bounding the agent SHALL NOT be certifiable.
APAS-CORE-10.3.2 X3, C4, and E3 operation, and any engagement whose affordance profile permits irreversible actions or autonomous chaining into adjacent surfaces, SHALL additionally require:
(a) continuous availability of the stop channel during execution;
(b) automated detection of scope-boundary approach or violation that halts or blocks the action rather than only recording it;
(c) declaration of how the authorized-scope boundary is enforced at runtime;
(d) halt-on-loss-of-oversight per 10.2.3;
(e) enforcement of the declared escalation limits (Section 7.1) by a mechanism whose failure is detectable;
(f) where actions are composed into a chain, the applicable gate SHALL be set at the worst reachable Action Reversibility Class across the chain before the first step executes.

10.4 Agent integrity

APAS-CORE-10.4.1 A candidate system SHALL implement and declare controls against adversarial manipulation of the agent itself during testing, covering at minimum (a) prompt injection from target content or responses; (b) manipulation or substitution of tools, plugins, models, or MCP and equivalent tool-connection servers; (c) poisoning of retrieved or persisted context; and (d) unauthorized modification or bypass of runtime controls or configurations governing scope, permitted tools or privileges.
APAS-CORE-10.4.2 Any Q3, Q4, R2, or R3 claim SHALL include in the quality declaration an agent-integrity disclosure of the controls in place, any manipulation attempts detected in the engagement, and their handling. A candidate-generated record of no detected manipulation SHALL NOT by itself constitute evidence that none occurred. A detected manipulation-class result meeting the Confirmed Finding bar under 4.3 SHALL enter the machine-readable finding export under 8.4.2; otherwise it remains quality-declaration content.
APAS-CORE-10.4.3 Undetected adversarial manipulation that materially altered a candidate system’s strategy, findings, scope adherence, or report content SHALL invalidate the affected claim dimensions. Agent-integrity evaluation procedures, including adversarial test cases and the limb (b) comparison method, SHALL be defined in the Evaluator Guide. That evaluation SHALL (a) extend at least as far as the autonomy class under 10.3.1 requires, and SHALL also scale with declared affordance risk and the reliance the claim carries under 10.4.4; and (b) compare labeled-evaluation against blinded runs on scope adherence, out-of-scope refusal, coverage, and finding-fabrication rate. Behavioral difference on those measures between labeled-evaluation and blinded runs is eval-awareness; material eval-awareness SHALL be disclosed and SHALL cap the assurance outcome.
APAS-CORE-10.4.4 Any Q3, Q4, R2, or R3 claim SHALL be supported by evaluator-executed adversarial agent-integrity testing covering at minimum the manipulation classes in 10.4.1, including goal hijack, tool-response injection, and context or memory poisoning. An agent-integrity disclosure of solely vendor-asserted controls SHALL NOT support a Q3, Q4, R2, or R3 claim.
APAS-CORE-10.4.5 On detection of adversarial manipulation, the candidate system SHALL suspend offensive activity, notify the target owner, the evaluator where present, and the emergency contacts under 7.4.3 within the declared maximum notification time, and SHALL quarantine evidence and findings from the suspected manipulation window for re-establishment.

10.5 Evaluation integrity

APAS-CORE-10.5.1 The truth set, evaluation environment, and scoring or result-submission channel SHALL be isolated from the candidate system’s reachable surface, with no read or write path to them from it.
APAS-CORE-10.5.2 Detected truth-set access by the candidate system, or a result record it submitted rather than one derived by the evaluator, SHALL invalidate the claim as an out-of-scope action does under 10.1.2.
APAS-CORE-10.5.3 The APAS Evaluator Guide SHALL define under 12.1.1 the procedure for establishing and verifying the 10.5.1 isolation. Where an excluded testing class is material to the profile, the Evaluator Guide SHOULD require the evaluator to seed at least one instance of it to test whether the exclusion is truthfully declared.

11. Public Claims and Marketing Language

11.1 Required disclosure

APAS-CORE-11.1.1 Any public APAS claim SHALL disclose:

  • atomic or matrix-based status
  • all eight claim dimensions
  • evaluator identity
  • certification and validity dates
  • candidate system version
  • material limitations where reliance is claimed
  • availability of the applicable quality declaration or a summary
  • where more than one combination is supported, the status of adjacent combinations a reasonable buyer might infer are available
  • the Core Standard version and the identifier and version of each companion document applied
  • where material, whether dynamic switching among combinations is supported
  • wherever the claim string is used publicly or in buyer-facing, marketing, or procurement-facing material, a one-sentence plain-language reading of the claim in the form of the Appendix A examples, with prominence equal to the encoded string

11.2 Prohibited claims

APAS-CORE-11.2.1 (a) No party SHALL state or imply, solely on the basis of APAS certification:

  • universal human equivalence or universal pentest substitution
  • framework-wide compliance or assessor acceptance by default
  • substitution without declared limitations and a represented subset of activities
  • higher pentest quality solely from higher execution autonomy
  • equivalence between testing modes absent explicit justification against a declared assurance objective
  • full-scope red-team equivalence on the basis of a T4 claim
  • platform-wide capability where only one combination is certified; implied certification of adjacent uncertified combinations; a capability matrix hiding unsupported, retired, failed, or unevaluated combinations
  • complete or full coverage absent a supporting coverage declaration, independent surface baseline where required, and methodology
  • coverage percentages or completeness figures whose denominator is solely the candidate system’s own discovery, where 4.2.4 applies
  • high assurance, assessor readiness, or substitution readiness solely because a report has few or no findings
  • benchmark leadership or public-corpus performance as proof of Q3, Q4, or R3
  • safety, fitness for deployment, or fitness to operate against any environment beyond the declared affordance profile and environment class (b) No party SHALL represent internal validation, multi-agent agreement, or consensus among the candidate system’s components as independent confirmation, evaluator confirmation, or EV3, or a non-rediscovery by its own exploration as absence, resolution, or non-existence, in any coverage, negative-result, retest, or public-claim statement (4.6.4). (c) No party SHALL represent or operate an APAS-certified system as an unrestricted, unauthorized, or general-purpose autonomous offensive capability, and APAS certification SHALL NOT be cited as endorsement of such use (2.2.1).

APAS-CORE-11.2.2 No party SHALL market an execution class, collaboration class, assurance outcome class, or reliance outcome without the full claim dimensions.
APAS-CORE-11.2.3 A Single-Capture Confirmation under 4.3.3 SHALL be labeled as such in any public claim and SHALL NOT be represented as independently reproduced.

12. Governance and Companion Documents

12.1 Companion documents

APAS-CORE-12.1.1 (a) The APAS document set SHALL comprise:

Companion document Purpose and scope
APAS Core Standard core claim semantics and minimum normative concepts
APAS Evaluator Guide run, scoring, and adjudication procedures; safety and stop-authority testing; agent-integrity adversarial testing; comparative validation; coverage-fidelity and independent-baseline construction; anti-gaming controls
APAS Profile Registry requirements for T1, T2A, T2C, T3, T4, and framework-specific profiles; mode-selection guidance; assumed-breach conditions for T4; context-rich testing requirements; any permitted carry-forward rules
APAS Attestation Schema machine-readable fields for claims, quality, coverage, evidence, and findings, profiled against the Section 8.4 interchange formats: combination identifiers, stable finding identifiers, observed operating path fields, capability-matrix statuses and a standardized capability-matrix template, evidence-manifest hash and signature requirements, and SARIF, OSCAL, and OpenVEX profiles or export mappings; the Agent Execution Record schema and semantic conventions under 8.4.7
APAS Crosswalk Annexes (optional) framework crosswalks, framework-specific reliance notes, additional reliance conditions, and limitations

(b) Where this Core document defers implementation detail, the applicable companion SHALL define the binding requirement. (c) Where it requires quality or coverage declarations, the Attestation Schema SHALL define the machine-readable fields, consistent with 8.4.
Note (informative). Three companions contemplated during drafting, an APAS Profile Runtime, an APAS Threat Model, and an APAS Certification Scheme Manual, are deferred to a later Core version. No requirement in this version depends on them: the Agent Execution Record schema is owned by the Attestation Schema under 8.4.7, the manipulation-class floor is stated in 10.4.1 and the testing procedure in 10.4.3, and scheme-operation rules presuppose the stewarding body required by 12.4.1.
APAS-CORE-12.1.2 No Atomic System Capability Certification or Engagement Conformance Statement SHALL be issued unless the Evaluator Guide, the applicable Profile Registry entry, and the Attestation Schema are published at compatible versions under 12.2.2.

12.2 Versioning

APAS-CORE-12.2.1 Major versions SHALL be used for changes to claim semantics, quality-dimension semantics, conformance structure, or any class or outcome definitions; minor versions SHOULD be used for new definitions, non-breaking requirements, or additional profile support; patch versions SHOULD be limited to errata and clarifications. These rules govern every APAS document.
Version classification (informative): v1.0 is major because it changes claim-validity semantics, quality-dimension floors, environment-class definitions, conformance publication gates, and finding and outcome semantics.
APAS-CORE-12.2.2 Each companion SHALL declare the Core major version against which it is normative. A claim SHALL be issued only against a companion set whose declared Core major version matches the attestation’s stated Core version. The stewarding body SHALL publish the compatible version set per Core major version.
APAS-CORE-12.2.3 (a) With each new Core major version, the stewarding body SHALL publish a transition period keeping the prior compatible set valid for claims already issued. (b) Publication of a new Core major version SHALL NOT retroactively invalidate a claim already issued against a previously compatible set; the claim remains valid on that set through its expiry or revalidation date under 5.4, with its superseded status disclosed, and SHALL state the Core major version and the compatible set it was issued against. (c) A claim SHALL NOT be issued against the prior compatible set after the transition period ends. (d) A change to this Core or a companion document is not a material change to the candidate system, and SHALL NOT by itself trigger 5.11.3. (e) A material candidate-system change under 5.11.3 requires re-attestation against the current compatible set.
APAS-CORE-12.2.4 The Core standard and Evaluator Guide SHALL be reviewed for capability drift on a defined cadence not exceeding twelve months, and SHALL define an out-of-cycle review trigger for material shifts in frontier autonomous-pentest capability.

12.3 Public review

APAS-CORE-12.3.1 Normative changes to the Core SHALL undergo public review before release of any version 1.0 or later. APAS SHALL be maintained through an open, versioned governance process with public comment tracking and public disposition of substantive comments.

12.4 Neutral stewardship

APAS-CORE-12.4.1 APAS SHALL be stewarded by a body organizationally independent of any single vendor operating in the autonomous pentesting market. A standard authored or controlled by a market participant SHALL NOT be represented as a neutral industry standard.
APAS-CORE-12.4.2 The stewarding body SHALL maintain a multi-stakeholder process in which no single vendor or group of affiliated vendors holds a controlling position over normative content. Participation SHOULD include vendors, end users, assessors, and independent researchers.
APAS-CORE-12.4.3 All participants in normative decisions SHALL disclose material conflicts of interest, including commercial interest in products the standard governs.

12.5 Status change, revocation, and dispute

APAS-CORE-12.5.1 The stewarding body SHALL maintain a public claim-status record. A claim’s status is Active, Under Review, Revoked, or Retired (S7).
APAS-CORE-12.5.2 On credible evidence that a claim rests on fabricated or materially misleading evidence (Section 8.2.4), the claim SHALL be moved to Under Review and SHALL NOT be relied upon as Q3, Q4, R2, or R3 until resolved.
APAS-CORE-12.5.3 A claimant or relying party MAY dispute a status change or an Unsupported-Finding downrate (4.3.2 and 4.3.8). The stewarding body SHALL define a time-bound adjudication process and SHALL publish the outcome and rationale.
APAS-CORE-12.5.4 A revoked claim SHALL be recorded with its effective date, and no party SHALL represent it as current.

13. Interpretation Rules

APAS-CORE-13.1 A profile MAY add stricter requirements but SHALL NOT weaken Core claim semantics, quality-dimension semantics, or safety requirements.
APAS-CORE-13.2 The attested declarations govern any conflicting public claim, and any public marketing summary conflicting with a quality or coverage declaration.
APAS-CORE-13.3 If a framework or assessor requires conditions stricter than APAS, those conditions govern for that use.
APAS-CORE-13.4 If execution class, collaboration class, assurance outcome, quality declaration, and reliance outcome appear to conflict, the conflict SHALL be resolved under 5.2.4 and 7.5.1; only where those do not resolve it do the evaluator result and attested declaration determine which dimension governs.
APAS-CORE-13.5 Where this Core document and a companion conflict, Core governs claim semantics, quality-dimension semantics, and safety requirements per 13.1, and the companion governs implementation detail expressly deferred under 12.1.1. Where two companions conflict, the Attestation Schema governs machine-readable field definitions and the Profile Registry profile-specific thresholds.

References

URLs accessed October 23, 2026.

Appendix A: Example Claim Declarations

Example claim Reading
APAS BB-T2A / X1-C1-Q2 / E1 [P-014] {R1} Human-led pentest with agent assistance; decision-useful output with declared coverage and limitations.
APAS WB-T3 / X2-C2-Q3 / E2 [P-031] {R2} Agent-led code-aware testing with active human collaboration; reference-validated output measured against an independent surface baseline.
APAS WB-T2C / X2-C2-Q4 / E3 [P-021] {R2} Agent-led testing with repeated human-agent handoff and rich implementation context; assessor-ready output for augmentation use.
APAS BB-T2C / X3-C4-Q4 / E3 [P-022] {R3} Independent agent execution with no in-band human collaboration; assessor-ready output for conditional substitution use. Requires disclosure of the threat-model rationale for BB, the resulting limitations, the independent surface baseline used, and the represented subset of activities. As an X3-C4 claim, it carries the strongest stop-authority, halt-on-loss-of-oversight, and agent-integrity obligations under Section 10.

Appendix B: Companion Documents

The companion document set is defined normatively in the 12.1 table; this appendix is a pointer only, keeping appendix lettering continuous with v0.9.

Appendix C: One-Sentence Core Standard

Full statement. An autonomous system may claim APAS-certified pentest capability only when, within a declared mode, tier, execution class, collaboration class, assurance outcome class, environment class, affordance profile, and reliance outcome, it produces bounded, authorized, integrity-protected pentest evidence whose coverage fidelity against an independent surface baseline, finding validity, evidence traceability, remediation usability, retestability, process integrity, and agent integrity can be evaluated by a relying reviewer within the limits of the applicable profile and attestation, with explicit disclosure of the context used, the materially tested and untested scope, and the capability combination actually evaluated.
Executive reading. An APAS claim states what an autonomous pentest system was shown to do, under what conditions, and how far a reviewer may rely on it; it holds only within the declared scope and conditions and only for the exact capability combination evaluated, and what was not tested, or could not be tested, must be stated.

Appendix D: Glossary

Finding states are defined in 4.3, quality dimensions in 4.1, and claim-dimension classes in Section 5. In this Core document:
Action Reversibility Class The most consequential effect an action may have on target state: read-only, externally reversible (by the target owner by a declared means), or irreversible (not reversible by the owner: destruction, one-shot consumption, irrevocable external effects).
Affordance Profile The declared operating envelope during testing: capabilities, restrictions, access conditions, escalation limits, and constraints, with the fields of 7.1.1.
Aggregation Rule The declared rule combining per-run outcomes into one claimed result, from the closed set in 4.6.5.
APAS Claim A bounded statement of what a candidate system demonstrated in an APAS evaluation, within a declared scope, conditions, and limitations.
Assessor-Consumable Evidence Package A package complete under 8.1.2, with the pentest report and traceability from findings and coverage statements to supporting artifacts, structured for reviewer evaluation of scope, coverage, findings, methodology, limitations, and the claimed reliance outcome without vendor-side analytical review.
Assurance Outcome The declared class (Q1 to Q4, 5.9) for the strength and usability of the pentest output.
Authorization Declaration The declaration required by 7.4.1.
Candidate System The autonomous product, service, or agentic workflow submitted for evaluation or used in an APAS-conformant engagement. A composite of cooperating agents is one candidate system for claim purposes; its components and orchestration are declared in the Agent Topology field (7.1.1).
Capability Combination A unique tuple of the claim-dimension classes (Section 5) with the affordance profile.
Capability Matrix A structured record of the capability combinations a candidate system version supports, has been evaluated for, and may publicly claim.
Confirmed Attack Path A finding state; see 4.3.7.
Coordinated-Disclosure Path The declared process for reporting and handling third-party, dependency, or out-of-scope vulnerabilities (7.4.2).
Coverage Declaration A structured record of tested, partially tested, untested, excluded, unreachable, and context-constrained scope against the applicable surface baseline, with the fields of 4.2.1.
Declared Methodology The documented methodology, scope model, coverage obligations, quality-dimension, evidence, and reporting requirements of the declared APAS profile.
Evaluator The party determining whether a candidate system or engagement satisfies applicable APAS requirements; independence is governed by 6.1.1.
Human Assistance Any human action materially affecting strategy, exploitation, validation, prioritization, control mapping, or report content. In-band analytical human assistance is such input during execution.
Identity Coverage The principals, roles, and tenants exercised during testing and the cross-principal and cross-tenant attempts (4.2.1), as distinct from those provisioned (7.1.1).
Independent Surface Baseline An enumeration of the in-scope attack surface from at least one source not under the candidate system’s sole control (owner inventory or evaluator-directed enumeration), against which its discovery and coverage are measured.
Materiality A fact, condition, finding, limitation, or discrepancy is material if its disclosure, omission, or misrepresentation could reasonably be expected to influence a relying reviewer’s assurance, reliance, or risk-acceptance decision on the pentest evidence, or would, if known, have led that reviewer to reject the evidence, lower the reliance outcome, or require additional testing. It governs disclosure and reporting, not material non-determinism, which has its own trigger in 4.6.1.
Negative-Result Statement Required by 4.5.1.
Observed Operating Path The sequence of materially relevant execution, collaboration, context, and assistance states actually used in a run.
Out-of-Band Scoring An evaluation result derived by the evaluator through a channel the candidate system can neither read from nor write to.
Out-of-Band Witness Evidence Evidence produced by an instrument the candidate system cannot author, observe, or reach: a callback listener, a canary, or a target-side monitor.
Pre-Registered Run Design The run count, selection rule, and purpose for runs informing a claim, fixed before execution and attested under 4.6.2.
Probabilistic Rediscovery A later run in which the candidate system finds, or fails to find, a previously reported condition through its own non-deterministic exploration rather than the recorded Reproduction Procedure.
Profile A normative APAS companion specification adding requirements for a specific testing objective, coverage model, or framework-aligned use case.
Quality Declaration A structured record accompanying a claim, recording the applicable quality dimensions, coverage declaration reference, finding-state summary, evidence references, remediation and retest status, and integrity disclosures (4.1.2).
Reachable Attack Surface The portion of the applicable surface baseline reachable under the declared affordance profile, environment class, credentials, vantage points, and context sources. Where no independent baseline is required, the declared in-scope surface lawfully reachable under those conditions (4.2.6).
Reference Validation Comparative validation against a declared reference method, cohort, or equivalent baseline under the same declared conditions.
Reliance Outcome The declared class (R1 to R3, 5.10) for the use a relying reviewer may make of the evidence package.
Relying Reviewer The auditor, assessor, or customer-appointed reviewer who decides whether an APAS evidence package suffices for the intended assurance or compliance use (6.2.1).
Reproduction Procedure A recorded, deterministic, agent-independent sequence of steps, requests, payloads, preconditions, and expected observations sufficient for a competent third party to reproduce a finding’s vulnerable condition without access to the candidate system. (a) Expected observations SHALL include the control or expected-secure observation obtained under the same conditions, or the procedure SHALL state why none was obtainable; (b) that alternative SHALL NOT apply to an authorization or tenancy-boundary finding, for which 4.4.6 requires paired observations; and (c) a procedure MAY be probabilistic only where the underlying condition is itself probabilistic (race condition, time-of-check to time-of-use defect, timing side channel, cache race), and SHALL then record attempt count, observed success rate, and conditions of measurement.
Single-Capture Confirmation A Confirmed Finding established under the 4.3.3 exception from preserved first-capture evidence.
Stable Finding Identifier A unique, persistent identifier assigned to a finding at first report and preserved across revisions, retests, exports, and format conversions.
Stop Authority The parties empowered to suspend or terminate execution (10.2.1), with the mechanism, channel, and maximum reaction time under 10.2.2.
Tamper-Evident Log An execution and evidence log kept so later insertion, deletion, or modification of entries is detectable, maintained or countersigned as 8.2.3 requires.
Truth Set The recorded ground-truth conditions present or planted in an evaluation environment, against which reported findings and coverage are scored; construction is defined in the APAS Evaluator Guide.
Unsupported-Finding Rate Unsupported findings as a proportion of all reported findings, reported under 4.3.8.
Vendor-Side Analytical Review Any vendor-originated interpretation, editing, prioritization, control mapping, or conclusion-shaping applied during or after execution, before delivery to the relying reviewer, including automated components that shape conclusions without re-observing the target.

Change Log

Version #                       Approved by Changes
20261023                       JY Original release
                           

Premier AI Safety Ambassadors

Premier AI Safety Ambassadors play a leading role in promoting AI safety within their organization, advocating for responsible AI practices and promoting pragmatic solutions to manage AI risks. Learn more about how your organization could participate and take a seat at the forefront of AI safety best practices.

Explore More of CSA

Research & Best Practices

Stay informed about the latest best practices, reports, and solutions in cloud security with CSA research.

Upcoming Events & Conferences

Stay connected with the cloud security community by attending local events, workshops, and global CSA conferences. Engage with industry leaders, gain new insights, and build valuable professional relationships—both virtually and in person.

Training & Certificates

Join the countless professionals who have selected CSA for their training and certification needs.

Industry News

Stay informed with the latest in cloud security news - visit our blog to keep your competitive edge sharp.