Skip to content

Security: google/cybernetic-agent-governance-engine

SECURITY.md

Security Policy

Supported Versions

Security fixes land exclusively on main; older release tags do not receive backports.

Version Supported
main (post-v4.1.0) ✅ Active (receives all security fixes)
4.1.0 ⚠️ Update to main (or the next release) for fixes published after v4.1.0
< 4.1.0 (4.0.x, 3.x, 2.x) ❌ No

Security Advisories

Published security advisories, affected ranges, CVSS vectors, and fix commits are maintained canonically on GitHub:

Every published GitHub Security Advisory has an in-repository fix-trace row below, mechanically verified against GitHub's security-advisories?state=published API and main git ancestry by scripts/check_advisory_traces.py:

Advisory Published Severity Fix on main Summary
GHSA-gc2v-v2jw-q499 2026-11-19 Medium 3feffdf6 (#820, client-v0.4.0) cage-client <= 0.3.0 and GatewayClient: HTTP 403 refusals and 202/409/422 deferrals accepted without GovernanceEnvelope signature verification
GHSA-c9j3-q7hh-p95m 2026-11-19 Medium dd27b66b (#815, AP-1 in docs/BREAKING_CHANGES.md) Governed financial advisor: CAGE_ALLOW_ASSERTED_PRINCIPAL accepted unverified end-user identities in enforcing postures (production/staging) despite documentation stating it was ignored there
GHSA-5c7p-g73w-v4m3 2026-11-19 Medium aa50d17d (#814) Governed inference proxy: non-content response fields bypassed output governance and PII redaction
GHSA-797p-rrgm-4p56 2026-11-19 High c145a9cc (#813) Governed financial advisor: require_api_key() accepted end-user OIDC tokens on service-only routes
GHSA-cxxx-6xfq-vvp4 2026-11-18 Medium f6313b20 (#809) provider_05 three-axiom attestation providers accepted live /v0.3/ responses as VERIFIED when v0.3 signature fields were stripped
GHSA-5438-3wwf-qvcq 2026-11-18 Medium 1ad49289 (#808) actuator_03 quarantine adapter settled unsigned or unverifiable receipts as ACCEPTED without ACTUATOR_03_REQUIRE_SIGNED_RECEIPTS=true
GHSA-348p-gfpg-qqm9 2026-11-18 Medium 1aeed422 (#799) provider_06 evidence adapter sealed receipts not bound to the submitted evidence and returned unverified seals without a key manifest
GHSA-c8pw-f3g3-xg7c 2026-10-26 Medium be23480b (v4.0.0) Finance causal tier risk boundary tripped 10,000x too high and zero-variance placebo refutation could not fail
GHSA-f6xq-23f7-9grm 2026-10-26 Medium 863d4092 (v3.3.0) EU_ECB FRIA revocation not re-checked on POST_HITL; execute_trade_bounded and BTC case variants bypassed trade_governance.rego
GHSA-w39v-5v7h-fxg9 2026-10-24 Medium cec7ab29 (#529, v3.3.0) Model-controlled trader_role raised OPA fiscal limits on the committing trade path

Earlier fixes resolved through public issues, internal audits, or pre-publication tracker entries are recorded in CHANGELOG.md and docs/POAM.md.

⚠️ Reference architecture notice: CAGE is a reference architecture and is not deployed to production. Advisories are published and fixed to ensure the codebase accurately represents the security posture claimed in associated research and compliance artifacts.

Reporting a Vulnerability

Please do not report security vulnerabilities through public GitHub issues.

To report a security vulnerability, please use the GitHub Security Advisory "Report a Vulnerability" feature. Please include:

  • A description of the vulnerability and its potential impact
  • Steps to reproduce the issue
  • Any proof-of-concept code (if applicable)
  • Your suggested fix (if you have one)

You should receive a response within 5 business days. If you do not receive a response, please follow up to ensure your report was received.

Disclosure Policy

We follow a coordinated disclosure model:

  1. You report the vulnerability privately.
  2. We confirm receipt and begin investigation within 5 business days.
  3. We develop and test a fix.
  4. We release the fix and publish a security advisory.
  5. You may publicly disclose the vulnerability after the fix is released, or after 90 days from the initial report — whichever comes first.

Landing an Advisory Fix (Maintainers)

When this flow applies vs. a direct public PR

  • Use the private-fork T=0 advisory flow below whenever a fix alters a security, authentication, authorization, cryptographic, PII-redaction, or governance-tier boundary in a distributed artifact (cage-client, cage-conformance, container images) or in main enforcement/policy code, and the PR diff itself exposes a real bypass or forgeability gap in prior commits or releases.
  • Use a direct public PR (no advisory) for hygiene cleanups, test-only or CI-gate additions, documentation/compliance updates, or structural refactors that close no bypass in any shipped artifact or main enforcement path. When uncertain whether an exposure warrants an advisory, default to the T=0 flow.

Procedure

GitHub Actions does not run on pull requests in temporary private forks, and ruleset 25966171 (Require status checks on main, no bypass actor) requires nine checks pinned to GitHub Actions — including Approved review or recorded bypass, whose post-merge review-bypass-ledger.yml check fails on main if a merge lacks a pre-merge review or Sigstore bypass record. Do not disable ruleset 25966171 or add a bypass actor. Ruleset 26377312 (blocking non-fast-forward updates and deletion) also stays enabled throughout.

Because pushing a fix branch to the public repository exposes the vulnerability path before publication, publish first at T=0 and land via normal public pull request(s):

  1. Develop and rehearse off-public: Stage fixes in the advisory's temporary private fork or a local isolated worktree. Before T=0, rebase onto current origin/main, run deterministic pre-commit checks and targeted security regression tests, and when landing multiple advisories together verify conflict-free multi-way merging across git merge-tree --write-tree (including shared files such as docs/BREAKING_CHANGES.md).
  2. Pre-stage draft advisory metadata via the REST API: Before T=0, PATCH each draft on repos/google/cybernetic-agent-governance-engine/security-advisories/<GHSA_ID> with its final public description (including a concrete non-code mitigation and with any <!-- ... --> comment templates removed), ecosystem: "other" structured range (patched_versions: null when remediating on main ahead of a tag), and verified CVSS 3.1 vector. Set reporter .credits once before acceptance, and omit credits from all subsequent PATCH payloads once credits_detailed[].state == "accepted" (passing credits again resets an accepted credit record back to pending). Confirm HTTP 200 on GET read-back so T=0 requires only a state transition.
  3. Close all open PRs in the temporary private fork before publishing: GitHub's advisory publish endpoint rejects {"state": "published"} with HTTP 422 while any pull request in the temporary private fork remains open. Close open private-fork PRs before switching the advisory state.
  4. At T=0, publish first, then open public PR(s):
    • PATCH each draft with {"state": "published"} only (leaving .credits and pre-staged metadata untouched).
    • Immediately push the pre-verified fix branch(es) to origin and open public pull request(s) against main, then PATCH each live advisory's remediation line to cite the open PR (Fix in flight in #<PR> alongside the immediate mitigation, omitting credits). While a newly published advisory is <= 24h old and carries Fix in flight in #<PR> (before Fixed on main in), scripts/check_advisory_traces.py (--verify-github) treats it as in-flight so #<PR>'s own CI stays green before its squash-merge commit SHA exists on main.
  5. Run all checks, squash-merge, stamp, and verify read-back:
    • Wait for all automated checks in ruleset 25966171 (strict up-to-date branch policy is disabled, so independent PRs can run checks in parallel). Never bypass a failing or pending check; rerun transient runner flakes, and if a PR stays stalled on CI past the disclosure window, update the in-flight advisory note to cite the staged PR commit SHA explicitly.
    • Squash-merge once checks pass (gh pr merge <PR> --squash --admin, where --admin in single-maintainer posture bypasses only ruleset 19769096's review rule while all nine checks in ruleset 25966171 pass).
    • Immediately stamp Fixed on main in <SHA> (#<PR>) onto the published advisory (description only — never pass credits), perform a GET read-back on the advisory endpoint to verify state == "published", the landed main SHA, and credits_detailed[].state == "accepted", and land the matching row in ## Security Advisories above so scripts/check_advisory_traces.py (--verify-github --strict) passes on main. Delete temporary branches and worktrees after post-merge checks pass on main.

Pre-remediated batch disclosures (fixes already on main)

When one or more draft advisories are held for a coordinated release date after their fix commits have already landed on main:

  1. Pre-stage the SECURITY.md trace branch locally before T=0: Add the new ## Security Advisories row(s) citing each already-merged main fix SHA and the scheduled publication date, and run uv run python scripts/check_advisory_traces.py --verbose locally to confirm every SHA is already an ancestor of origin/main (do not open the PR before publishing, or --verify-github will reject unpublished GHSA IDs).
  2. Publish and open the trace PR back-to-back at T=0: PATCH each draft advisory with {"state": "published"} only (never passing credits), verify state == "published" and credits_detailed[].state == "accepted" via GET read-back, and immediately push and open the pre-staged SECURITY.md PR. On that PR's CI run, the API's newly published set matches SECURITY.md 1-to-1 with zero missing rows, keeping advisory-trace-parity green throughout the disclosure.

Package Distribution Channels

cage-client and cage-conformance are distributed exclusively via GitHub Releases (client-v* / conformance-v* tags with attached wheels/sdists) and GitHub Packages (GHCR) — not PyPI. PYPI_TRUSTED_PUBLISHING_ENABLED in .github/workflows/publish-client-sdk.yml is intentionally left unset/disabled due to Google branding and package-publishing policy and must not be enabled without explicit branding sign-off. Security advisories therefore use ecosystem: "other" with commit-SHA / GitHub Release remediation, and scripts/check_client_sdk_sync.py enforces [tool.cage.release] parity against GitHub release tags.

Scope

The following are in scope for security reports:

  • Remote code execution in the governance gateway or compliance bridge
  • Authentication/authorisation bypass in the governance pipeline
  • Governance tier bypass (violations of the NoDirectBind invariant)
  • Injection vulnerabilities (prompt injection, SQL injection, etc.)
  • Cryptographic weaknesses in the Cloud KMS signing, kid-resolved trust anchors, development-only software signers (Ed25519 / HMAC-SHA256), routing seals, or the SHA-256 hash-chain implementation
  • Control Barrier Function (CBF) race conditions or invariant violations
  • Secrets or credentials exposed in the repository

The following are out of scope:

  • Vulnerabilities in third-party dependencies (report these upstream)
  • Denial-of-service attacks requiring physical access
  • Social engineering attacks
  • Issues in documentation only

Security Hardening Notes

CAGE is a domain-agnostic governance substrate designed for regulated environments generally; finance (src/cage_finance/) is the sole finalized, production-eligible domain plugin (healthcare in src/cage_healthcare/ and physical_ai in src/cage_physical_ai/ ship as non-production reference stubs; POAM-2026-076–078, POAM-2026-198), and the jurisdictional region (US_FED, EU_ECB or APAC_MAS, set by CAGE_DEPLOYMENT_REGION and read only by resolve_deployment_region(); unknown values are refused, and an unset value is refused in enforcing postures and resolves to US_FED only in dev, test and ci) determines which regulatory controls are in scope for a given deployment. Key security controls are documented in:

Implemented Controls Summary

Control Census Claim ID Implementation
Governance signing sp800-53-component-definition.yaml#61c43f8a-6693-4049-ae09-edd8b22f500a Cloud KMS asymmetric signing (cloud providers in src/integrations/{gcp,aws,azure}/kms_provider.py, loaded via signer_factory.py); separate keys per role (gateway seal, reconciler snapshot, compliance-bridge evidence) with kid-resolved verification. Software Ed25519 / HMAC signers are refused under an enforcing posture and allowed only in dev/test/CI. 90-day rotation cadence per KEY_ROTATION.md
Routing seal (v3) sp800-53-component-definition.yaml#0c2d40d7-0d19-4b9d-9afa-6c80967b31b4 Single-use JWT signed by the gateway's governance signer (Cloud KMS in enforcing postures), verified by kid against an independently resolved trust anchor and bound to the SHA-256 evidence record_hash. The HMAC v2 4-tuple is a legacy path verify_seal() rejects while CAGE_SEAL_STRICT_MODE is true (the default) and always in production. Internal to /tools/execute and ConsequenceGateway — not used for caller authentication (partially-implemented: hermetic round-trip verified in test_routing_seal_asymmetric_roundtrip.py; strict-mode and live KMS round-trip read-back open under POAM-2026-235)
TLS & Transport Security sp800-53-component-definition.yaml#230de322-4b76-4a0d-b07d-267f4fa23769 Redis TLS 1.2+ enforcement (src/gateway/infrastructure/redis_client.py), OIDC JWKS verify=True enforcement, and Linkerd mTLS across meshed pods (tests/test_tls_enforcement.py). Outbound HTTP client wiring for create_hardened_client_context() (src/gateway/infrastructure/tls_context.py) is open under reopened POAM-2026-011 (docs/POAM.md)
Base Image Hardening system-security-plan.yaml#s3000001-plat-si3-8000-platform0001 Advisor, gateway, compliance-bridge and NeMo Guardrails images use two-stage builds on a digest-pinned Wolfi base (cgr.dev/chainguard/wolfi-base) with no compiler, util-linux or package-manager cache in the runtime stage and a build-time apk upgrade; images are tagged by git SHA and their digests signed for Binary Authorization at build time (scripts/build_images.sh) (lula-validation-si2.yaml active; SI-3 runtime admission stub open)
Prompt injection detection sp800-53-component-definition.yaml#8e46d991-dd78-416a-897d-f909cb63be8d Aho-Corasick O(n) scan; 14+ patterns (partially-implemented: Stage 1 deterministic scan active; semantic tier opt-in)
PII protection component-definition.yaml#e1f2a3b4-c5d6-7890-efab-901234567890 Presidio; 15 entity types; input + output (POAM-2026-146, POAM-2026-147 open)
Human-in-the-loop component-definition.yaml#aarm2026-v7-context-overflow Redis-persisted checkpoint; TOCTOU remediation via post_hitl_rehydrate + post_hitl_revalidate (advisor calls the gateway's POST /governance/revalidate-post-hitl) (POAM-2026-196 open)
Control Barrier Function sp800-53-component-definition.yaml#943496c8-5b07-48e4-a175-731366d416fa Atomic Redis Lua (atomic_verify_and_commit()) with synchronous replica WAIT barrier, monotonic safety:fence_epoch with a shared high-water mark (safety:fence_epoch_hwm), and fail-closed state rollback
Evidence chain integrity sp800-53-component-definition.yaml#80d95b3a-12ae-4dd4-8db4-b52661335fa4 SHA-256 hash-chained NDJSON & Redis Streams db=1 (noeviction, Lua compare-and-append); enforced blocking durability in production (validate_evidence_stream_preconditions()); compliance-bridge EvidenceCustodian KMS batch attestation (EVIDENCE_KMS_KEY) to WORM storage and CustodyVerifier read-back verification (kid-resolved trust anchors, assert_citable() OSCAL citation gate)
mTLS sp800-53-component-definition.yaml#5f386745-cdda-431b-adec-850cdd899f3e Linkerd mTLS with a Google CAS trust anchor and defaultInboundPolicy = "all-authenticated" (infra/modules/service_mesh); per-workload Linkerd Server / HTTPRoute / AuthorizationPolicy across gateway, compliance-bridge, :9090 metrics, and :4191 admin ports (POAM-2026-007, POAM-2026-080, POAM-2026-195 closed 2026-11-13; component-definition census WEAKER pending restatement)
Caller identity sp800-53-component-definition.yaml#ca8f96e2-69a3-4d6b-980c-962e73e075d8 Gateway ingress requires a Linkerd-verified l5d-client-id that matches CAGE_TRUSTED_CLIENT_IDENTITIES (WorkloadIdentityMiddleware); enforced in every environment, no header/body identity fallback, fail-closed 403. An RFC 9449 DPoPValidator (src/gateway/server/dpop_validator.py) ships and is unit-tested but is not yet wired into any ingress path. See docs/architecture/AGENT_IDENTITY_BINDING_SPEC.md (partially-implemented: SVID binding enforced; DPoP unwired)
Egress credentials (v3.1.0, seam only — POAM-2026-072) system-security-plan.yaml#h8000009-5000-4000-8000-000000000017 CredentialBrokerAdapter Protocol in Layer 1 (src/gateway/governance/seams/credential_broker.py); when a concrete broker is supplied to Actuator01BindingAdapter, credentials are fetched per dispatch, masked in logs, excluded from ActuationReceipt, and fail closed on CredentialNotFound / CredentialAccessDenied via ingest_actuation_receipt(). At HEAD no concrete CredentialBrokerAdapter ships in src/ and BrokerActuator reads BROKER_API_KEY from os.environ (POAM-2026-072 reopened)
Egress lockdown system-security-plan.yaml#p7000001-plat-sc7-8000-platform0001 GKE Dataplane V2 NetworkPolicy & FQDNNetworkPolicy resources in infra/targets/gcp-gke/network_policy.tf (POAM-2026-190, POAM-2026-195 closed; L3/L4 FQDN wildcard scope open under POAM-2026-152)
Token quota enforcement component-definition.yaml#tqp007-req-0001-0000-0000-000000000001 Per-session step-count (≤12) and token (≤100k) via Redis atomic Lua counters; fail-closed (partially-implemented: Lua quota enforced; lula-validation-tqp007.yaml is rated Stub: it checks only that the gateway is ready with a REDIS_URL, not fail-closed behaviour, POAM-2026-219)

Note: CAGE is a reference architecture. Regulated-environment deployers must conduct their own risk assessment before production use. For current status see docs/POAM.md and the security and compliance status section of the root README.md.

KMS Cryptographic Signing Security

Replay-attack closure: KMS-signed reconciliation payloads embed a signed_at Unix timestamp. The verifier (KMSGovernanceSigner.verify() in kms_signer.py) rejects any payload where now - signed_at > 300 s (MAX_KMS_PAYLOAD_AGE_SECONDS).

Redis / Data-Layer Security

Strict Replication & Fence Epoch Hardening: In production (CAGE_ENV=prod), ControlBarrierFunction asserts synchronous replica replication (CAGE_STRICT_REPLICATION=true). If replica synchronization fails during WAIT, the local balance debit is automatically rolled back (rollback_state(cost)) and fails closed. Monotonic safety:fence_epoch counters prevent stale-replica balance replays.

Evidence Stream Precondition Hardening & Custody Verification

Audit Durability & Citation Guarantee: validate_evidence_stream_preconditions() is posture-based: under an enforcing posture (anything but dev/test/ci; an unset CAGE_ENV is production) it halts startup if EVIDENCE_STREAM_ENABLED=false, or if EVIDENCE_CHAIN_BLOCKING=false without an explicit CAGE_ALLOW_NONBLOCKING_PROD=true. The gateway lifespan also starts the sink via start_evidence_sink() and fails closed, so no routing seal or ConsequenceGateway EXECUTE verdict is issued without durable evidence commitment to the tamper-evident log. In the compliance bridge, EvidenceCustodian signs batches with EVIDENCE_KMS_KEY and writes them via put_if_absent() to WORM cold storage, while CustodyVerifier (evidence_verifier.py) re-verifies the archive on EVIDENCE_VERIFY_INTERVAL_S against kid-resolved trust anchors and gates OSCAL assessment citations (OSCAL_REQUIRE_VERIFIED_CUSTODY=true).

Zero-Trust Caller Identity

Header-spoofing closure: Caller identity is never read from application-layer data. X-Agent-ID headers, body-supplied agent_id fields, and the X-CAGE-Routing-Seal header are not accepted as authentication. The sole source of truth is the l5d-client-id header, which the Linkerd inbound proxy sets from the verified mTLS peer certificate (or strips when there is none). The gateway accepts exactly one such header matching CAGE_TRUSTED_CLIENT_IDENTITIES, in every environment; anything else fails closed with 403 rather than falling back to an anonymous principal. The mesh AuthorizationPolicy independently admits only the advisor's service account (POAM-2026-080 & POAM-2026-167, closed in certified Gate W 8956f505).

Egress Credential Brokerage Seam (v3.1.0 — POAM-2026-072 Open)

Seam contract (concrete broker open under POAM-2026-072): CredentialBrokerAdapter defines the Layer 1 per-dispatch credential seam: when a concrete broker is injected into Actuator01BindingAdapter, credentials are requested per dispatch keyed on the calling agent's SVID and tool name, injected as request headers only in local scope, and excluded from ActuationReceipt, findings, and HTTP response bodies (test_credential_headers_not_in_audit_record). At HEAD zero concrete CredentialBrokerAdapter implementations ship in src/, Actuator01BindingAdapter skips the broker step when broker=None, and BrokerActuator reads BROKER_API_KEY from os.environ (reopened POAM-2026-072 in docs/POAM.md).

Masking is prefix-preserving, not total. Log masking applies value[:8] + "****". For an Authorization: Bearer <token> header only the scheme survives, but a broker that returns a bare-token header will leak the first 8 characters of the secret into logs. Brokers should return scheme-prefixed header values.

Prohibited Security Anti-Patterns

The following patterns are strictly forbidden in production code and will be rejected in review:

1. Hardcoded Cryptographic Secrets (CWE-798)

  • PROHIBITED: secret = os.environ.get("API_KEY", "default-secret")
  • REQUIRED: Fail-closed validation asserting os.environ["API_KEY"] exists and meets minimum entropy length (>= 32 chars).

2. Shell Injection Vectors (CWE-78)

  • PROHIBITED: subprocess.run(command, shell=True)
  • REQUIRED: subprocess.run(shlex.split(command)) with direct argument lists.

3. TLS/SSL Verification Bypass (CWE-295)

  • PROHIBITED: requests.get(url, verify=False)
  • REQUIRED: Explicit TLS certificate verification (verify=True).

4. Insecure Deserialization (CWE-502)

  • PROHIBITED: pickle.loads(data) or unconstrained yaml.load(data)
  • REQUIRED: json.loads(data) or yaml.safe_load(data).

Enforcement

All prohibited patterns are enforced via static grep checks, pre-commit hooks, CI security jobs, and targeted unit regression suites in tests/test_*_security.py.

Security Research Acknowledgments

The CAGE maintainers thank the following researchers for reporting security and governance boundary observations with reproducible test cases:

  • Daniel Fairbanks (@danofairanks) — Boundary analyses and hardening reports across CAGE's runtime governance, provenance, and offline evaluation planes (#379, #405, #406, #463, #464, #465, #472, #473, #511, #512, #513, #514, #516), including trajectory NARROW vs. ALLOW metric separation, multi-turn shadow-state tier parity, RefusalReceipt proof_hash verification and DPO chosen-argument sanitization, offline drawback-detector, SKILL.md static-linter, rollout-runner symbol/HTTP bounds, fail-closed Stage 2 certification gating (--stage0-only), positive-lift and drawback enforcement in run_certification_gate(), STPAValidator action normalization / approval-token validation / core-rule preservation, actuator_02 signed-receipt enforcement, FTRA NaN-confidence guards, FTRA registry expires_at enforcement, retry-safe evidence stream appends and chain-truncation detection, stage- and evidence-bound .cer skill-certification receipts (cage.skill.cer.v2), full-order routing-verdict reuse and multi-step trade plan previews, MCP send-phase retry boundaries and ExecutionLedger trade idempotency, and explicit end-user principal assertion boundaries.
  • @moiznationzz — Reported the compliance-bridge dual-control quorum bypass via unverified x-cage-source-principal headers, CAGE_ENV default authentication bypass in de5534ac:deployment/k8s/compliance-bridge.yaml, and uninvoked verify_webauthn_challenge() in DeferQueue.approve() (GHSA-529c-23c8-3m42).
  • @Zorta-Zenith — Independently reported the x-cage-source-principal multi-party quorum forgery and OIDC fallback gap in src/compliance_bridge/auth.py (GHSA-qrg7-27x4-mq39).
  • @m0ssser — Reported the docker-compose.yml entrypoint override exposing mcp_tool_server.py's POST /tools/execute (trigger_safety_intervention) without WorkloadIdentityMiddleware (GHSA-4x3x-w2m3-8mfh).
  • @ArksherX — Reported Control Barrier Function (CBF) ground-truth balance verification and atomic reservation boundary findings (GHSA-cmwm-295c-gjf6, GHSA-w7g7-wjj9-6j97, GHSA-w7f4-8grh-2xv8).
  • @uwezkhan — Reported the inference proxy multi-role / streaming governance bypass (GHSA-hfqj-24cj-693g) and /governance/validate-action ingress authentication boundary (GHSA-v3h4-8458-5ww3).
  • Dominik Blain (@dom-omg), Ironproof — Reported, with step-by-step reproductions, that a normative-provider determination revoked during human review was never re-checked after approval (the EU_ECB fria tier did not run under POST_HITL), that execute_trade_bounded and case or whitespace variants of BTC bypassed trade_governance.rego, and that the FRIA payload named no caller and the provider_01 request relied on self-asserted mandate facts (GHSA-f6xq-23f7-9grm, POAM-2026-113). Thank you for a careful, clearly reasoned report.
  • @kitkatt7 — Reported and submitted remediations for partner adapter verification boundaries across provider_06 evidence-hash binding (GHSA-348p-gfpg-qqm9, #799), actuator_03 signed quarantine receipt enforcement (GHSA-5438-3wwf-qvcq, #808), and provider_05 live /v0.3/ attestation signatures (GHSA-cxxx-6xfq-vvp4, #809), as well as earlier public hardening submissions (#517; #662, consolidated into #742).
  • @katesaikishore — Reported that non-content OpenAI-compatible response fields (reasoning aliases, refusal, audio, annotations, logprobs, legacy function_call, and raw SSE replay) bypassed governed inference proxy output governance and PII redaction (GHSA-5c7p-g73w-v4m3, fixed via explicit egress-allowlist response reconstruction in #814).
  • @oyesanyf — Reported that require_api_key() in the governed financial advisor accepted end-user OIDC tokens on service-only routes (GHSA-797p-rrgm-4p56, fixed in #813).
Learn more about advisories related to google/cybernetic-agent-governance-engine in the GitHub Advisory Database