AAES Govern checks permission before a registered tool action proceeds on an enforced path. Permitted capabilities, task-scoped grants, human approval and budget ceilings can limit what an agent is allowed to do, including when an attacker has influenced its instructions. Sealed records support later review of the decisions AAES recorded.
The mapping below uses the OWASP Top 10 for Agentic Applications 2026. It is AAES's assessment of its own controls. OWASP has not reviewed or endorsed this mapping. The LLM Top 10 note covers the separate model-application risk list.
These pages describe the product. They are not a certification or a legal opinion.
Reading this note
Each risk has a relevant control or a stated gap. A control's relevance does not mean it resolves the whole risk. Enforcement requires AAES Govern to control the action's credential path; observation alone cannot stop an action. A permitted tool can still produce an unsafe result.
Use Action controls for control preconditions and Offline verification for evidence limits. The OWASP publications define the risks; the control-to-risk relationships here are AAES's interpretation.
Instrument identity
- Title
- OWASP Top 10 for Agentic Applications 2026. The Agent Control Standard (ACS) is a separate specification: specification v0.1.0, project release 0.1.4 (16 October 2026). ACS-Core and ACS-Trace change between releases of the same specification version.
- Issuing body
- OWASP Foundation (community project)
- Instrument type
- Community security-risk and mitigation guidance. ACS separately defines runtime-control requirements and conformance profiles. These publications are not regulations.
- Official sources
- Agentic Top 10 2026 publication, OWASP's risk-category reference, and the Agent Control Standard repository.
The agentic track is the part of the GenAI Security Project concerned with systems that take actions, not only systems that generate text. Its threat publications provide mitigation guidance. ACS v0.1.0 also defines normative runtime-control requirements and conformance profiles, including a mandatory ACS-Core baseline. The project publishes schemas and reference implementations. Conformance claims are self-declared, and no conformance test suite ships yet.
Question
Which tool actions remain possible if the agent follows a malicious instruction?
Start with the capabilities the agent can call, the task and lifetime of its grants, and the approvals required before execution. On a brokered path, AAES uses the configured credential to execute a permitted call without handing that credential to the agent. Supported spending requests also face a budget decision before dispatch. These controls can constrain actions while the model's instructions remain compromised.
Review the tool's own permissions and behavior too. Permission to call a tool does not establish that its implementation is safe, that its output is trustworthy, or that its downstream actions are governed.
ASI01 to ASI10 mapping
The risk identifiers refer to the 2026 Agentic Top 10. OWASP's current category reference calls ASI08 "Cascading Agent Failures"; the publication's launch article calls it "Cascading Failures".
| OWASP risk | Relevant AAES controls | Remaining gap and client responsibility |
|---|---|---|
| ASI01: Agent Goal Hijack Tests and evidence limits | Permitted capabilities, scoped grants and named approval apply to routed actions regardless of the instruction that prompted them. | AAES does not detect or remove prompt injection. A hijacked agent can still take actions the policy permits. Protect input and retrieval sources and review permitted actions. |
| ASI02: Tool Misuse & Exploitation Tests and evidence limits | Capability permissions and task scope restrict routed tool use. Actions classified as irreversible require an authorized person's approval. Budget ceilings refuse supported spending requests above the remaining ceiling before dispatch. | A permitted call can still misuse a tool. Review action classifications, request contracts and server-side authorization. Argument values are constrained only where the registered contract checks them; budgets limit authorized amounts, not final provider charges. |
| ASI03: Identity & Privilege Abuse Tests and evidence limits | Task-scoped grants have a finite lifetime. On brokered paths, the configured credential stays with the executor. Records attribute the decision to the actor, task and grant. | Alternate credentials can bypass AAES. Remove those paths and apply least privilege at the tool and identity provider. Recorded attribution alone does not prove identity-provider security. |
| ASI04: Agentic Supply Chain Vulnerabilities Tests and evidence limits | Capability permissions limit which registered tools an agent can invoke through AAES. Brokered calls take their destination and credential from configured bindings. | This does not establish tool or dependency integrity. Review and maintain models, packages, MCP servers and tool implementations. A compromised registered tool can act maliciously after an authorized call. |
| ASI05: Unexpected Code Execution Tests and evidence limits | Where code execution is a registered capability routed through AAES, capability permissions, task scope and applicable approval rules govern access to that capability. | AAES authorization does not establish code safety or provide a general sandbox. Isolate execution, restrict network and filesystem access, and validate code inputs. Execution outside AAES bypasses these checks. |
| ASI06: Memory & Context Poisoning Tests and evidence limits | Governed actions still face capability, scope and approval checks when the agent's memory or context is poisoned. Sealed records support review of journaled decisions. | AAES does not validate or repair agent memory, embeddings or retrieved content. Protect their write paths and provenance. A valid decision record does not prove that its inputs were true. |
| ASI07: Insecure Inter-Agent Communication Tests and evidence limits | A receiving agent's registered tool actions can have their own capability, grant and approval checks when those actions route through AAES. | This is a downstream action control. It does not establish authentication, encryption, replay protection or message validation for arbitrary agent-to-agent traffic. Secure each communication channel separately. |
| ASI08: Cascading Agent Failures Tests and evidence limits | Task scope and supported-spend ceilings can limit the actions AAES authorizes. A journal write failure prevents new grants. | These controls do not establish workflow recovery or stop every failure from spreading. Journal failure alone does not revoke existing grants; their issued lifetimes are bounded by 15 minutes. Evaluate dependencies, retry behavior and recovery separately. |
| ASI09: Human-Agent Trust Exploitation Tests and evidence limits | Agents cannot approve themselves. Named human approvals bind to a specific intent; actions classified as irreversible require an authorized person. | A person can approve a misleading or harmful request. Give reviewers enough context to assess the action and its consequences. Approval attribution establishes who approved, not whether the judgment was sound. |
| ASI10: Rogue Agents Tests and evidence limits | Capability permissions, finite grant lifetimes and approval rules constrain governed actions. Sealed records let reviewers check the integrity of the exported decision history. | AAES does not infer an agent's hidden goals or prove that it behaved safely. A rogue agent may misuse permitted actions or bypass AAES with other authority. Review access and investigate activity using additional telemetry. |
Control, test and evidence links
The local shared evaluation suite 3.0.0 records 123/123 passing checks across LangGraph, CrewAI and Microsoft AutoGen AgentChat/Core. Each framework runs 26 native calls, 3 scripted approval checks, 10 fixture controls and 2 fault-injection probes. These are selected control checks; none is a direct test of an entire OWASP risk category.
The machine-readable evidence map identifies the case IDs, result-object pointers, artifact hashes, record hashes and sidecar/export line numbers. Its artifact paths are relative to the local external package. The framework evidence download page lists the published versioned packages, checksums and candidate qualification. The map is a separate AAES-authored interpretation; the original evaluation signatures do not sign the OWASP relationships.
- Local package
external-evaluation-package-v5-20261016, public candidateexternal-v3-candidate.5- Package index SHA-256
c5d734d0bed1be9a136a4d73cb033b0513eef419826f068efe7e6db9fd358837- Source snapshot commitment
38faa343ea0a9729f3cf2103c104738cd6402a8167b4b1d8e06285171203f790. Proprietary source and producer binaries are omitted from the external package; this commitment does not verify their contents there.- Supplied evidence
- 81 records re-sealed and bound to three exports. Each history is CONSISTENT with no independent witness or TSA.
| OWASP risk | Control evidence | Untested mechanism or remaining limit |
|---|---|---|
| ASI01: Agent Goal Hijack | No prompt-injection or goal-hijack attack is executed. These are related action-control checks. | |
| ASI02: Tool Misuse & Exploitation | Selected tool-action boundaries are tested. Tool implementation safety and arbitrary argument validation remain outside these cases. | |
| ASI03: Identity & Privilege Abuse | The foreign-task case is a tenant/task boundary check. It does not validate an identity provider, credential custody or grant expiry. | |
| ASI04: Agentic Supply Chain Vulnerabilities | No model, dependency, MCP-server or tool supply-chain integrity test is included. | |
| ASI05: Unexpected Code Execution | No code-execution payload or sandbox escape is tested. Authorization checks do not establish code safety. | |
| ASI06: Memory & Context Poisoning | No memory, retrieval or context poisoning is tested. A consistent record does not prove truthful inputs. | |
| ASI07: Insecure Inter-Agent Communication | No inter-agent authentication, encryption, message validation or channel replay test is included. Effect-key retry checks do not establish channel security. | |
| ASI08: Cascading Agent Failures | The kit tests selected local retry, restart and journal Record API refusal/recovery behavior. Distributed failure propagation, physical storage outages and revocation of existing grants are not evaluated. | |
| ASI09: Human-Agent Trust Exploitation | The operator is scripted. No human persuasion, reviewer comprehension or approval-interface test is included. | |
| ASI10: Rogue Agents | The kit does not infer hidden goals, exercise bypass credentials or establish safe agent behavior. |
Configured action permissions
Selected synthetic cases. The selected read is allowed and the configured destructive action is refused before simulated dispatch. Control preconditions.
Shared cases: read_allowed, policy_refused.
- These cases do not test prompt-injection detection, arbitrary tool argument contracts or malicious tool implementations.
Intent-bound scripted approval
Selected synthetic cases. The selected spend requires approval, refuses self-approval and changed payloads, and rejects an expired approval. Control preconditions.
Shared cases: approval_required, agent_self_approval_refused, scripted_operator_approval, changed_payload_refused, approved_action_allowed, expiry_approval_required, scripted_approval_expiry, expired_operator_approval_refused, expired_approval_cannot_authorize.
- Operator decisions and clock advances are scripted. The kit does not demonstrate human review, an operator UI or the quality of a reviewer's judgment.
Task ownership, cancellation and deadline
Selected synthetic cases. The selected calls refuse a task owned by another tenant, a cancelled task and a task past its deadline. Control preconditions.
Shared cases: cross_tenant_task_refused, cancellation_baseline_allowed, scripted_task_cancellation, cancelled_task_refused, scripted_deadline_expiry, expired_task_refused.
- This tests task checks on the local brokered path. It does not establish identity-provider security, finite grant lifetime enforcement or downstream credential revocation.
Task commitment ceiling
Selected synthetic cases. The selected task permits its first commitment and refuses further commitments above its remaining ceiling, including after a restart. Control preconditions.
Shared cases: budget_first_allowed, budget_exhausted_refused, budget_after_restart_refused.
- The ceiling applies to task commitments in the simulated workflow. Actual provider charges, other credentials and spending outside AAES are not evaluated.
Duplicate and uncertain retry protection
Selected synthetic cases. The selected duplicate and uncertain calls cannot execute again under the same effect key, including after a helper restart or crash. Control preconditions.
Shared cases: duplicate_retry_refused, uncertain_effect_preserved, uncertain_retry_refused, concurrent_duplicate_refused, scripted_process_restart, duplicate_after_restart_refused, uncertain_after_restart_refused, crash_after_simulated_effect, crash_retry_after_restart_refused.
- The race is a scripted Go-handler probe. The crash exits the helper after persisting a simulated effect and before returning its outcome. Native framework parallelism, distributed concurrency, PostgreSQL failover and live-provider reconciliation are not evaluated.
Supplied export and record integrity
Supplied-artifact verification. The public candidate authenticates disclosed artifacts, verifies the exports, re-seals current records and binds their projections. Control preconditions.
Public test references: TestRecordResealingAndBindingMutations in verifier-source/internal/evaluation/records_test.go, line 51. TestSignedExportMustMatchSealedProjection in verifier-source/internal/evaluation/records_test.go, line 147. The package qualification records the module tests passing; it has no separate per-function receipt.
The machine map links the recorded package qualification checks, including export/record mutation refusals, wrong-key refusal and the required-independent-attestation refusal.
- The histories have no independent witness or TSA. Checks do not establish complete capture, truthful inputs, correct decisions or real effects. Receipt-chain seals and omitted proprietary source are not verified. Legacy record seals, retention tombstones and nonempty external-reference sets are refused.
Downstream credential custody and grant lifetime
Not evaluated in this kit. The expanded kit supplies no direct test of downstream credential custody or grant lifetime enforcement. Control preconditions.
- The fixture uses a synthetic local credential and executor. TCP/TLS, remote identity, real credential handling, alternate credentials and provider-side revocation remain outside the kit.
Refusal when the decision journal cannot be written
Selected synthetic cases. The selected native calls are refused with no returned grant or simulated dispatch when Record rejects before append, returns saturation, or errors after durable append. Retrying the same intent after recovery executes once. A durable allowed record can remain even though the failed request did not dispatch. Control preconditions.
Shared cases: scripted_journal_record_failure, journal_record_failure_refused, scripted_journal_record_recovery, journal_record_retry_after_recovery, scripted_journal_saturation_failure, journal_saturation_failure_refused, scripted_journal_saturation_recovery, journal_saturation_retry_after_recovery, scripted_journal_commit_failure, journal_commit_failure_refused, scripted_journal_commit_recovery, journal_commit_retry_after_recovery.
- Errors are injected by an evaluation-only wrapper at the production journal Record API. This does not induce physical disk/fsync failures, exercise PostgreSQL failover, test receipt-write failure, prove revocation of existing grants or establish production outage behavior. Pre-append refusals have no new journal record; the error after durable append retains one allowed record without a provider effect.
Shared case index
Skip to verification and deployment limits
Show all 41 cases
Journal refusal cases inject pre-append errors, saturation and errors after durable append at the journal Record API. A failed call must return no grant and make no simulated provider call. The recovery retry uses the same intent. These cases do not induce physical storage faults or revoke existing grants.
Case kinds distinguish native framework execution from scripted approvals, fixture controls and Go-handler probes. A case can have no new journal record; its signed result still records the scripted observation. Exact references for all three frameworks are in the machine map.
| Case ID | Kind | Expected observation | Recorded results |
|---|---|---|---|
read_allowed | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
policy_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
approval_required | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
agent_self_approval_refused | approval | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_operator_approval | approval | granted | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
changed_payload_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
approved_action_allowed | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
duplicate_retry_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
uncertain_effect_preserved | runtime | uncertain | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
uncertain_retry_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
budget_first_allowed | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
budget_exhausted_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
cross_tenant_task_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
cancellation_baseline_allowed | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_task_cancellation | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
cancelled_task_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
expiry_approval_required | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_approval_expiry | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
expired_operator_approval_refused | approval | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
expired_approval_cannot_authorize | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_deadline_expiry | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
expired_task_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
concurrent_duplicate_refused | probe | one_allowed_one_refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_process_restart | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
duplicate_after_restart_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
uncertain_after_restart_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
budget_after_restart_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
crash_after_simulated_effect | probe | crashed_after_effect | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
crash_retry_after_restart_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_record_failure | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_record_failure_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_record_recovery | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_record_retry_after_recovery | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_saturation_failure | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_saturation_failure_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_saturation_recovery | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_saturation_retry_after_recovery | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_commit_failure | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_commit_failure_refused | runtime | refused | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
scripted_journal_commit_recovery | control | completed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
journal_commit_retry_after_recovery | runtime | allowed | autogen: recorded PASS crewai: recorded PASS langgraph: recorded PASS |
Simulator metadata
The manifest and configuration labels agree in each of these three histories. The machine map includes both signed field references for each framework.
- autogen: manifest
aaes.local-provider/2; configurationaaes.local-provider/2. - crewai: manifest
aaes.local-provider/2; configurationaaes.local-provider/2. - langgraph: manifest
aaes.local-provider/2; configurationaaes.local-provider/2.
Verification and deployment limits
- This is AAES's mapping of selected controls to the existing OWASP note. It is not an OWASP review, endorsement, risk-coverage score or ACS conformance result.
- All provider effects are local simulations. Passing cases do not establish model quality, real effects or production readiness.
- Approvals, clock advances and fault-injection operations are scripted. The race and crash probes are not native framework parallel execution.
- All three histories are CONSISTENT with no independent witness or TSA. Bundled keys establish consistency; signer identity requires separately trusted keys.
- Proprietary source and producer binaries are omitted. Their matching signed commitments do not verify the omitted contents.
- Receipt-chain seals are not verified. All disclosed receipt bytes are authenticated by the original evaluation manifest.
- The candidate re-seals supported current record/2 records. Legacy seals, retention tombstones and nonempty external-reference sets are refused.
- Task checks do not establish downstream credential revocation. PostgreSQL failover, distributed concurrency, live-provider reconciliation and other crash boundaries are outside the suite.
- Native package execution qualification covers macOS ARM64 only.
- The mapping is a separate AAES-authored interpretation, checksummed without a new digital signature. The original evaluation signatures authenticate their own artifacts, not these OWASP relationships.
Preconditions and gaps
- ACS conformance remains unverified for AAES. ACS v0.1.0 has normative requirements and no conformance test suite yet. AAES has not published a requirement-by-requirement assessment. AAES also has a contribution proposal in the ACS public discussion (issue 185).
- Enforcement requires control of the credential path. Activity outside AAES Govern is visible only where telemetry is supplied. Observation is not enforcement.
- Offline integrity verification does not establish complete action history, decision correctness or real-world execution. The local v3 candidate used by the mapped evaluation kit also re-seals supported current records and binds them to export entries. Its partial profile leaves receipt-chain seals and omitted source unverified. This candidate is separate from the published v0.2.0 and v0.3.0 verifier releases.
- No production adoption as of this review. Deployments are self-hosted or hosted by AAES.
For implementation and testing details beyond this note, use Security and implementation posture and the evaluation page. Results from a client's test tenant are produced during a scoped evaluation.
Client responsibility
The organization runs its own security program against the project's guidance. The client remains the regulated entity.
During a scoped evaluation, use a test tenant to check the following:
- Try a permitted call and a call outside the agent's capability or task scope. Verify the decision and whether the tool was invoked.
- Request an action classified as irreversible without approval, then with approval for a different intent. Check that neither authorizes execution.
- Submit a supported spending request above the remaining budget. Check refusal before dispatch and inspect the recorded decision.
- Make the decision journal unavailable. Check the daemon's refusal to issue a new grant and account for grants already issued.
- Export the records and verify them offline. Use the records sidecar to check stored scope and approval attribution, and compare provider-side records when assessing external outcomes.
These are evaluation steps to perform in your deployment. They do not report a completed customer test or an independent assessment.
Choose the tool actions to test, identify who holds their credentials, and list the questions the decision records should answer.