Skip to content

armgit / AiSOC

Public Imported from GitHub
0

Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.

The one unpatchable dependency is removed, and the rest were already fixed (#791)

Beenu Aroracb55b439126c
diff --git a/.github/workflows/check-openapi.yml b/.github/workflows/check-openapi.yml
index e2ef2dd58..d76ad1ddb 100644
--- a/.github/workflows/check-openapi.yml
+++ b/.github/workflows/check-openapi.yml
@@ -19,7 +19,7 @@ env:
OPENAPI_DEPS: >-
fastapi uvicorn pydantic "pydantic[email]" pydantic-settings
"sqlalchemy[asyncio]" asyncpg structlog
- "python-jose[cryptography]" "passlib[bcrypt]"
+ PyJWT cryptography "passlib[bcrypt]"
tenacity pyyaml prometheus-client
neo4j redis celery aiofiles aiokafka
boto3 clickhouse-driver
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 5536fda4f..cb579fc29 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -397,7 +397,7 @@ jobs:
env:
API_DEPS: >-
fastapi "pydantic[email]" pydantic-settings "sqlalchemy[asyncio]" asyncpg
- structlog "python-jose[cryptography]" "passlib[bcrypt]" tenacity pyyaml
+ structlog PyJWT cryptography "passlib[bcrypt]" tenacity pyyaml
prometheus-client "opentelemetry-sdk" "opentelemetry-api"
"opentelemetry-exporter-otlp-proto-grpc"
"opentelemetry-instrumentation-fastapi"
diff --git a/.github/workflows/cross-tenant-rbac.yml b/.github/workflows/cross-tenant-rbac.yml
index 948aecb89..8de42cd8a 100644
--- a/.github/workflows/cross-tenant-rbac.yml
+++ b/.github/workflows/cross-tenant-rbac.yml
@@ -80,7 +80,7 @@ jobs:
# having to maintain a parallel pin list here.
API_DEPS: >-
fastapi "pydantic[email]" pydantic-settings "sqlalchemy[asyncio]" asyncpg
- structlog "python-jose[cryptography]" "passlib[bcrypt]" tenacity pyyaml
+ structlog PyJWT cryptography "passlib[bcrypt]" tenacity pyyaml
prometheus-client "opentelemetry-sdk" "opentelemetry-api"
"opentelemetry-exporter-otlp-proto-grpc"
"opentelemetry-instrumentation-fastapi"
diff --git a/.github/workflows/integration.yml b/.github/workflows/integration.yml
index 0890714f6..83be86500 100644
--- a/.github/workflows/integration.yml
+++ b/.github/workflows/integration.yml
@@ -295,7 +295,7 @@ jobs:
run: >-
pip install --quiet
fastapi "pydantic[email]" pydantic-settings "sqlalchemy[asyncio]" asyncpg
- structlog "python-jose[cryptography]" "passlib[bcrypt]" tenacity pyyaml
+ structlog PyJWT cryptography "passlib[bcrypt]" tenacity pyyaml
prometheus-client "opentelemetry-sdk" "opentelemetry-api"
"opentelemetry-exporter-otlp-proto-grpc"
"opentelemetry-instrumentation-fastapi"
@@ -587,7 +587,7 @@ jobs:
run: >-
pip install --quiet
fastapi "pydantic[email]" pydantic-settings "sqlalchemy[asyncio]" asyncpg
- structlog "python-jose[cryptography]" "passlib[bcrypt]" tenacity pyyaml
+ structlog PyJWT cryptography "passlib[bcrypt]" tenacity pyyaml
prometheus-client "opentelemetry-sdk" "opentelemetry-api"
"opentelemetry-exporter-otlp-proto-grpc"
"opentelemetry-instrumentation-fastapi"
diff --git a/.github/workflows/wet-eval.yml b/.github/workflows/wet-eval.yml
index 05172389f..84f78754f 100644
--- a/.github/workflows/wet-eval.yml
+++ b/.github/workflows/wet-eval.yml
@@ -121,7 +121,7 @@ jobs:
pip install --quiet \
"fastapi" "pydantic[email]" "pydantic-settings" \
"sqlalchemy[asyncio]" "asyncpg" "structlog" \
- "python-jose[cryptography]" "passlib[bcrypt]" \
+ PyJWT cryptography "passlib[bcrypt]" \
"tenacity" "pyyaml" "prometheus-client" \
"neo4j" "redis" "celery" "httpx" "aiofiles" \
"email-validator" "PyJWT" "aiosqlite" \
diff --git a/AGENTS.md b/AGENTS.md
index 1282ed7f1..955602009 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -23,7 +23,7 @@
- Monorepo managed with pnpm ([email protected]) and Turborepo; workspaces defined in `apps/*` and `packages/*`.
- Apps: `apps/web` (Next.js frontend), `apps/docs` (documentation site).
- Backend services in `services/`: `api` (FastAPI/Python 3.11), `agents`, `alert-fusion`, `connectors`, `demo-producer`, `enrichment`, `fusion`, `ingest`, `realtime`, `threatintel`, `ocsf`.
-- API service stack: FastAPI, Uvicorn, SQLAlchemy (async), asyncpg (PostgreSQL), Alembic (migrations), Redis, python-jose (JWT), Pydantic v2.
+- API service stack: FastAPI, Uvicorn, SQLAlchemy (async), asyncpg (PostgreSQL), Alembic (migrations), Redis, PyJWT (JWT), Pydantic v2.
- Packages: `packages/types` (shared TypeScript types), `packages/ui`, `packages/sdk-go`, `packages/sdk-py`, `packages/sdk-ts`, `packages/plugin-sdk-go`, `packages/plugin-sdk-py`.
- Docker Compose used for local dev: `infra/compose/` is the canonical home for all non-root compose files (`docker-compose.dev.yml`, `docker-compose.demo.yml`, `docker-compose.airgap.yml`); only the slim `docker-compose.yml` remains at repo root. Build contexts inside the moved files use `../../services/<name>` (two levels up). Terraform in `infra/terraform/` for infrastructure.
- CI uses GitHub Actions (`.github/workflows/`); includes workflows for OpenAPI checks, CI, docs deployment, marketplace sync, and detection validation.
diff --git a/CHANGELOG.md b/CHANGELOG.md
index ac463837b..1a60eec75 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -54,6 +54,45 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
The invite is single-use. The child-scoped routes answer `404` rather than
`403` for a tenant that is not yours, so they cannot enumerate tenant UUIDs.
+- **`python-jose` is gone, and `ecdsa` with it.** `ecdsa` carried
+ CVE-2024-23342 (Minerva timing attack on P-256) with no patched release —
+ OSV records the affected range as introduced at 0 with no fixed event,
+ because upstream states python-ecdsa offers no side-channel resistance and
+ will not fix it. It was an unconditional requirement of `python-jose`, so
+ the suppression was renewed rather than resolved. `services/api` now signs
+ and verifies with `PyJWT`, which it already depended on for the OIDC and
+ SAML paths; `python-jose`, `ecdsa` and `rsa` all leave the dependency tree.
+ Six CI workflows installed `python-jose[cryptography]` and never installed
+ `PyJWT`, and reached `cryptography` — a declared direct dependency of
+ `services/api` — only through that extra; they now install both by name.
+
+- **`image-size` moved to a patched release instead of staying suppressed.**
+ Both advisories were held open on the reading that no fix existed. They
+ record a vulnerable range of `<= 2.0.2`, and npm has published 2.0.3 and
+ 2.0.4; a null `first_patched_version` is not the same claim as no fix
+ existing. A pnpm override pins `>=2.0.4 <3`, which
+ `@docusaurus/mdx-loader`'s `^2.0.2` range accepts. This clears the only two
+ high-severity advisories in the pnpm workspace.
+
+- **The dependency suppression list is empty.** All 42 entries in
+ `scripts/security_audit_ignores.txt` were re-verified against OSV and
+ against the versions the lockfiles actually resolve. Every one was
+ resolvable, and most of the justifications had stopped being true: nine
+ starlette entries blamed a `fastapi<0.137` cap that exists in neither bot
+ service, three cryptography entries blamed `<50`/`<49` caps that exist
+ nowhere in the repository, and the langchain, weasyprint, anyio, aiohttp,
+ h2, idna and pydantic-settings entries each named a version older than the
+ one their lock resolves. The file now records what was measured, so the next
+ review starts from evidence rather than from the previous reason string.
+
+- **The pnpm and Go arms of the audit could report success without scanning.**
+ A failed `govulncheck`, an unparseable `pnpm audit` response and a registry
+ that never answered were all recorded as warnings, which exit 0 — so the job
+ printed "0 findings" for ecosystems it had not read. All three now record a
+ coverage gap, which `exit_code_for` already failed on for the Python arm.
+ That property — an unscanned target fails the build — had no test; it does
+ now.
+
### Fixed
- **Scheduled hunts ran against credentials that could not exist, so every one
diff --git a/apps/docs/docs/compliance/fips-posture.md b/apps/docs/docs/compliance/fips-posture.md
index fab2a8825..4791bca0c 100644
--- a/apps/docs/docs/compliance/fips-posture.md
+++ b/apps/docs/docs/compliance/fips-posture.md
@@ -18,7 +18,7 @@ of the answers.
| Credential encryption (`vault:v2`) | AES-128-CBC per-secret DEK, KEK in AWS KMS | `cryptography` + KMS | KMS side: yes |
| Backup encryption | AES-256-GCM | `cryptography` / OpenSSL | No |
| Audit hash chain | SHA-256 | `hashlib` / OpenSSL | No |
-| Session tokens | HMAC-SHA256 (JWT) | `python-jose` | No |
+| Session tokens | HMAC-SHA256 (JWT) | `PyJWT` / `hashlib` | No |
| Sandbox determinism | BLAKE2b | `hashlib` | **Not a FIPS algorithm** |
| TLS | Whatever the deployment terminates with | Operator's choice | Operator's choice |
@@ -45,7 +45,7 @@ Not a configuration flag. In rough order of effort:
incidental use — including BLAKE2b above, and any transitive dependency
reaching for MD5 in a non-security context, which several libraries do
for cache keys.
-3. **An audit of transitive dependencies.** `python-jose`, `passlib`,
+3. **An audit of transitive dependencies.** `PyJWT`, `passlib`,
`neo4j`, `clickhouse-driver` and the Kafka clients all perform
cryptographic operations. Each needs checking against the validated
module rather than assumed.
diff --git a/package.json b/package.json
index 69c9f7ee9..7e5bc5f24 100644
--- a/package.json
+++ b/package.json
@@ -82,6 +82,7 @@
"brace-expansion@1": ">=1.1.18 <2",
"browserslist": ">=4.28.7 <5",
"fast-uri": ">=3.1.6 <4",
+ "image-size": ">=2.0.4 <3",
"ip-address": ">=10.3.1 <11",
"js-yaml@3": ">=3.15.2 <4",
"js-yaml@4": ">=4.3.2 <5",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 050f33651..867856547 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -25,6 +25,7 @@ overrides:
brace-expansion@1: '>=1.1.18 <2'
browserslist: '>=4.28.7 <5'
fast-uri: '>=3.1.6 <4'
+ image-size: '>=2.0.4 <3'
ip-address: '>=10.3.1 <11'
js-yaml@3: '>=3.15.2 <4'
js-yaml@4: '>=4.3.2 <5'
@@ -3004,7 +3005,7 @@ packages:
estree-util-value-to-estree: 3.5.0
file-loader: 6.2.0([email protected])
fs-extra: 11.3.4
- image-size: 2.0.2
+ image-size: 2.0.4
mdast-util-mdx: 3.0.0
mdast-util-to-string: 4.0.0
react: 19.3.0
@@ -12570,9 +12571,9 @@ packages:
engines: {node: '>= 4'}
dev: true
- resolution: {integrity: sha512-IRqXKlaXwgSMAMtpNzZa1ZAe8m+Sa1770Dhk8VkSsP9LS+iHD62Zd8FQKs8fbPiagBE7BzoFX23cxFnwshpV6w==}
- engines: {node: '>=16.x'}
+ resolution: {integrity: sha512-QRUkFFsRV/6fuESxb9Vkq+a0LkSrgKXuc2NEqfikiXxxN/G3tjWt5EVUlMaImRBZRZK/jRBEbYvpPYZL8t08Zw==}
+ engines: {node: '>=18'}
hasBin: true
dev: false
diff --git a/scripts/security_audit.py b/scripts/security_audit.py
index 1471051bb..9f28351f4 100644
--- a/scripts/security_audit.py
+++ b/scripts/security_audit.py
@@ -212,13 +212,17 @@ def run_pnpm_audit(repo_root: Path, ignores: list[Ignore]) -> Report:
except json.JSONDecodeError:
print(f"Failed to parse pnpm audit output:\n{output}", file=sys.stderr)
report = Report()
- report.warnings.append("pnpm audit returned unparseable output")
+ # Unparseable output means the workspace was not audited. Recorded
+ # as a coverage gap rather than a warning for the same reason the
+ # Python arm does it: a warning still exits 0, so the job would
+ # report "pnpm: 0 findings" having scanned nothing at all.
+ report.unscanned.append("pnpm workspace: audit returned unparseable output — NOT scanned")
return report
return classify_pnpm_audit(data, ignores)
report = Report()
- report.warnings.append("pnpm audit failed on all registries")
+ report.unscanned.append("pnpm workspace: audit failed on every registry — NOT scanned")
return report
@@ -472,7 +476,14 @@ def run_govulncheck(repo_root: Path, ignores: list[Ignore]) -> Report:
break
if proc.returncode not in (0, 3):
- combined.warnings.append(f"govulncheck on {module} exited {proc.returncode}: {proc.stderr[:200]}")
+ # govulncheck exits 0 (clean) or 3 (vulnerabilities found); anything
+ # else — a build failure, a missing toolchain, an unresolvable
+ # module — means this module was not analysed. That is a coverage
+ # gap, not a warning: a warning exits 0 and the arm would print
+ # "go: 0 findings" for a module it never managed to read.
+ combined.unscanned.append(
+ f"{module}: govulncheck exited {proc.returncode} — NOT scanned ({proc.stderr[:160].strip()})"
+ )
continue
vulns = parse_govulncheck_json(proc.stdout)
diff --git a/scripts/security_audit_ignores.txt b/scripts/security_audit_ignores.txt
index 5d87a9d60..2c5c3f616 100644
--- a/scripts/security_audit_ignores.txt
+++ b/scripts/security_audit_ignores.txt
@@ -9,140 +9,85 @@
# Example:
# python|CVE-2026-12345|Upstream fix not released; tracked in issue #123|2026-07-15
-# ── 2026-09-21 re-triage ─────────────────────────────────────────────────────
-# The 2026-05-28 batch expired on 2026-08-25 and the 2026-06-27 batch was four
-# days from expiry, so both were re-checked against OSV and against the
-# versions the lockfiles actually resolve — not against the claims in the
-# superseded reason strings, several of which had gone stale (e.g.
-# "serialize-javascript 6.0.2 is latest" and "fastapi<0.137 cap in
-# services/api" were both no longer true).
-#
-# Anything with a reachable patched release was FIXED rather than re-ignored:
-# • pnpm.overrides in the root package.json now pin patched versions of
-# serialize-javascript, fast-uri, @babel/plugin-transform-modules-systemjs,
-# browserslist, js-yaml, svgo, sharp, brace-expansion, ip-address and
-# websocket-driver; apps/web moved to next 16.3.4. That took pnpm from 66
-# advisories / 28 high+critical to 34 / 2, and retired 23 of the 25 pnpm
-# suppressions below.
-# • services/connectors, services/fusion, services/threatintel and
-# services/osquery-tls still carried caret pins (fastapi ^0.109.0 /
-# ^0.111.0) that held them on starlette 0.36.3. They now use the same
-# >=0.109,<0.142 range as api/agents/actions, which resolves a clean
-# starlette, so CVE-2024-47874 (fixed 0.40.0) and CVE-2025-54121
-# (fixed 0.47.2) were dropped rather than suppressed.
-#
-# What remains is the set with no reachable fix: either upstream has published
-# none, or the fix sits behind a major migration that needs its own PR.
-
-# starlette — services/slack-bot and services/teams-bot still cap fastapi
-# <0.137, which resolves fastapi 0.136.1 / starlette 1.0.0. services/api,
-# services/agents and services/actions have already moved to <0.142 and
-# resolve starlette 1.3.1, which is clean. Clearing these needs the two bot
-# services off the <0.137 cap, which moves their Bolt / Bot-Framework adapter
-# surface, so it is tracked as a dedicated PR rather than an audit unblock.
-# Both CVE and GHSA forms are listed because pip-audit may report either as
-# the primary id depending on the advisory source.
-python|PYSEC-2026-161|starlette Host-header path poisoning (fixed 1.0.1); slack-bot/teams-bot resolve 1.0.0 under their fastapi<0.137 cap; tracked for the bot-service fastapi bump|2026-12-19
-python|CVE-2026-48817|starlette HTTPEndpoint dispatch via getattr (fixed 1.1.0); slack-bot/teams-bot resolve 1.0.0 under their fastapi<0.137 cap; tracked for the bot-service fastapi bump|2026-12-19
-python|GHSA-x746-7m8f-x49c|starlette HTTPEndpoint dispatch via getattr (fixed 1.1.0); slack-bot/teams-bot resolve 1.0.0 under their fastapi<0.137 cap; tracked for the bot-service fastapi bump|2026-12-19
-python|CVE-2026-48818|starlette SSRF via UNC paths on Windows (fixed 1.1.0); Linux-only deployment; slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137|2026-12-19
-python|GHSA-wqp7-x3pw-xc5r|starlette SSRF via UNC paths on Windows (fixed 1.1.0); Linux-only deployment; slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137|2026-12-19
-python|CVE-2026-54282|starlette hostname poisoning (low; fixed 1.3.0); slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137; tracked for the bot-service fastapi bump|2026-12-19
-python|GHSA-jp82-jpqv-5vv3|starlette hostname poisoning (low; fixed 1.3.0); slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137; tracked for the bot-service fastapi bump|2026-12-19
-python|CVE-2026-54283|starlette urlencoded form DoS (fixed 1.3.1); slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137; tracked for the bot-service fastapi bump|2026-12-19
-python|GHSA-82w8-qh3p-5jfq|starlette urlencoded form DoS (fixed 1.3.1); slack-bot/teams-bot capped at starlette 1.0.0 by fastapi<0.137; tracked for the bot-service fastapi bump|2026-12-19
-
-# idna — CVE-2026-45409 is fixed in 3.15. services/agents (3.18) and
-# services/teams-bot (3.15) are already clean; services/slack-bot still
-# resolves 3.13 through its own lock. It is a leaf transitive dep, so the fix
-# is a plain lock refresh on that one service — tracked, not a code change.
-python|CVE-2026-45409|idna punycode DoS (fixed 3.15); only services/slack-bot still resolves 3.13; leaf transitive dep, tracked for a slack-bot lock refresh|2026-12-19
-
-# langchain / langgraph — every one of these is fixed only in a release line
-# that our pins cannot reach: langchain-openai is capped <0.4.0 (fix landed in
-# 1.1.14), langchain <0.4.0 (fix in 1.3.9), langgraph <1.1.0 (langgraph-sdk fix
-# in 0.3.15 and langgraph-checkpoint fix in 4.1.1 both sit behind the langgraph
-# 4.x major). Those are breaking-API migrations that need their own PR with
-# regression coverage; folding them into a dependency-audit unblock is exactly
-# the kind of change that should not ride along unreviewed.
-python|PYSEC-2026-76|langchain-openai 0.3.35 (fixed 1.1.14); unreachable under the <0.4.0 cap, needs the langchain 0.3 to 1.x migration; tracked for a dedicated PR|2026-12-19
-python|PYSEC-2026-77|langchain-text-splitters 0.3.11 (fixed 1.1.2); transitive under langchain <0.4.0, needs the langchain 0.3 to 1.x migration; tracked for a dedicated PR|2026-12-19
-python|PYSEC-2026-83|langgraph 1.0.1 (fixed 1.0.10); reachable inside the <1.1.0 cap but the agents lock needs a coordinated langgraph refresh; tracked for a dedicated PR|2026-12-19
-python|CVE-2026-27794|langgraph-checkpoint 3.0.1 (fixed 4.0.0); requires the langgraph 4.x major bump; tracked for a dedicated PR|2026-12-19
-python|GHSA-gr75-jv2w-4656|langchain path traversal in file-search middleware (fixed 1.3.9); requires the langchain 0.3 to 1.3 major bump; tracked for a dedicated PR|2026-12-19
-python|CVE-2026-48775|langgraph-checkpoint unsafe JSON deserialization (fixed 4.1.1); requires the langgraph 4.x major bump; tracked for a dedicated PR|2026-12-19
-python|GHSA-fjqc-hq36-qh5p|langgraph-checkpoint unsafe JSON deserialization (fixed 4.1.1); requires the langgraph 4.x major bump; tracked for a dedicated PR|2026-12-19
-python|CVE-2026-48776|langgraph-sdk unsafe URL path construction (fixed 0.3.15); transitive via langgraph, blocked by the same major bump as CVE-2026-48775|2026-12-19
-python|GHSA-w39p-vh2g-g8g5|langgraph-sdk unsafe URL path construction (fixed 0.3.15); transitive via langgraph, blocked by the same major bump as CVE-2026-48775|2026-12-19
-
-# weasyprint / anyio — both come from services/agents, which is one of the three
-# services carrying a committed poetry.lock, so widening the pyproject range does
-# not move them: `poetry export` replays the lock. weasyprint 70.0 and anyio
-# 4.15.1 both exist and both fix these, so the remediation is a lock refresh on
-# that one service, which dependabot already covers via the pip/services/agents
-# entry in .github/dependabot.yml. Deliberately not hand-editing the lock.
-# The api service (no lock) is handled by widening its weasyprint pin in #636.
-# Both CVE and GHSA forms are listed since pip-audit may report either.
-python|PYSEC-2026-3940|weasyprint 69.0 SSRF (fixed 70.0); pinned by the committed services/agents lock, needs a lock refresh; api is covered separately by the <71 pin bump|2026-12-19
-python|CVE-2026-55073|weasyprint 69.0 SSRF (fixed 70.0); pinned by the committed services/agents lock, needs a lock refresh; api is covered separately by the <71 pin bump|2026-12-19
-python|GHSA-jf6q-chmf-3h3v|weasyprint 69.0 SSRF (fixed 70.0); pinned by the committed services/agents lock, needs a lock refresh; api is covered separately by the <71 pin bump|2026-12-19
-python|CVE-2026-63349|anyio 4.14.1 run_process ignores extra_groups (fixed 4.15.1); pinned by the committed services/agents lock, needs a lock refresh|2026-12-19
-python|GHSA-3w57-8xmc-8v26|anyio 4.14.1 run_process ignores extra_groups (fixed 4.15.1); pinned by the committed services/agents lock, needs a lock refresh|2026-12-19
-python|CVE-2026-64847|anyio 4.14.1 process-pool workers block on undrained stderr (fixed 4.15.1); pinned by the committed services/agents lock, needs a lock refresh|2026-12-19
-python|GHSA-5p39-cfhj-2xmp|anyio 4.14.1 process-pool workers block on undrained stderr (fixed 4.15.1); pinned by the committed services/agents lock, needs a lock refresh|2026-12-19
-python|CVE-2026-63374|anyio 4.14.1 advisory co-resolving with CVE-2026-63349/64847 (fixed 4.15.1); pinned by the committed services/agents lock, needs a lock refresh|2026-12-19
-
-# pydantic-settings — fixed in 2.14.2. services/agents already resolves 2.14.2;
-# services/slack-bot and services/teams-bot still resolve 2.14.1 through their
-# own locks. Patch-level bump, waiting on a lock refresh for those two.
-python|GHSA-4xgf-cpjx-pc3j|pydantic-settings symlink follow (fixed 2.14.2); agents is clean, slack-bot/teams-bot still resolve 2.14.1; tracked for a lock refresh|2026-12-19
-
-# ── 2026-07-11 triage (expire 2026-10-05) ────────────────────────────────────
-# ecdsa 0.19.2 is the latest published release (2026-03-26) and already carries
-# the CVE-2026-33936 DER-parsing fix. PYSEC-2026-1325 still flags 0.19.2, so
-# there is no newer release to upgrade to. ecdsa is a hard transitive dependency
-# of python-jose[cryptography] (pulled in regardless of the crypto extra) used by
-# the API JWT layer; it is never used to parse untrusted attacker-supplied DER
-# private keys. Removing it requires migrating the auth layer off python-jose to
-# PyJWT (already a dep) — a regression-prone change tracked for a dedicated PR.
-python|PYSEC-2026-1325|ecdsa 0.19.2 is latest release; no patched version available; transitive via python-jose[cryptography]; tracked for python-jose to PyJWT migration|2026-10-05
-
-# ── 2026-07-13 triage (expire 2026-10-11) ────────────────────────────────────
-# langchain-core prompt-loading path traversal (langchain_core.prompts.loading
-# reads files from paths in deserialized config dicts). Fixed only in
-# langchain-core 1.2.22 — a major bump from our pinned 0.3.x line with breaking
-# API changes, blocked by the same langchain 0.3 → 1.x migration tracked above
-# (GHSA-gr75-jv2w-4656 et al.). The vulnerable prompts.loading module is not
-# imported anywhere in this codebase (verified via repo-wide grep).
-# Both PYSEC and GHSA forms listed since pip-audit may report either as primary.
-python|PYSEC-2026-2193|langchain-core prompts.loading path traversal (fixed in 1.2.22); requires langchain 0.3 → 1.x major bump, module unused in codebase; tracked with langchain migration PR|2026-10-11
-python|GHSA-qh6h-p6c9-ff54|langchain-core prompts.loading path traversal (fixed in 1.2.22); requires langchain 0.3 → 1.x major bump, module unused in codebase; tracked with langchain migration PR|2026-10-11
-
-# langchain-core SSRF via image_url token counting in
-# ChatOpenAI.get_num_tokens_from_messages (low severity, freshly disclosed).
-# Fixed only in langchain-core 1.2.11 — same langchain 0.3 → 1.x major bump as
-# the entries above. The affected method is not called anywhere in this
-# codebase (verified via repo-wide grep for get_num_tokens_from_messages /
-# image_url), so no untrusted image URL ever reaches it.
-python|PYSEC-2026-2562|langchain-core SSRF in ChatOpenAI.get_num_tokens_from_messages (low; fixed in 1.2.11); requires langchain 0.3 → 1.x major bump, method unused in codebase; tracked with langchain migration PR|2026-10-11
-python|GHSA-2g6r-c272-w58r|langchain-core SSRF in ChatOpenAI.get_num_tokens_from_messages (low; fixed in 1.2.11); requires langchain 0.3 → 1.x major bump, method unused in codebase; tracked with langchain migration PR|2026-10-11
-
-# ── 2026-07-30 / 2026-08-04 triage (expire 2026-10-28) ───────────────────────
-# Python-side advisories held by version caps that need cross-service compat
-# testing. The pnpm entries that used to live in this block were retired by the
-# 2026-09-21 override pass above.
-python|CVE-2026-69247|cryptography high (48.0.1/49.0.0); fixed in 50.0.0 but held by the <50 and <49 version caps across services (49 to 50 is a major bump needing cross-service compat testing); tracked for a dedicated dependency PR|2026-10-28
-python|CVE-2026-69248|cryptography high (48.0.1); fixed in 50.0.0, held by the <50/<49 major-version caps; tracked for a dedicated dependency PR|2026-10-28
-python|CVE-2026-69249|cryptography high (48.0.1); fixed in 50.0.0, held by the <50/<49 major-version caps; tracked for a dedicated dependency PR|2026-10-28
-python|CVE-2026-59881|aiohttp high (3.14.1); fixed in 3.14.3 within the <4.0 cap; patch bump tracked for the next poetry lock refresh|2026-10-28
-python|CVE-2026-69243|aiohttp high (3.14.1); fixed in 3.14.3 within the <4.0 cap; patch bump tracked for the next poetry lock refresh|2026-10-28
-python|CVE-2026-69244|aiohttp high (3.14.1); fixed in 3.14.3 within the <4.0 cap; patch bump tracked for the next poetry lock refresh|2026-10-28
-
-# ── 2026-08-11 triage (expire 2026-11-05) ────────────────────────────────────
-# image-size is the one pnpm advisory pair with genuinely no fix: the GitHub
-# advisory records patched_versions as "<0.0.0", i.e. no released version is
-# clean, so an override has nothing to point at. Build-time only (the next
-# image pipeline), never fed untrusted input at runtime.
-pnpm|GHSA-5p2g-fcmc-qvqq|image-size 2.0.2; advisory reports no patched release (patched_versions "<0.0.0"), so an override has no clean target; build-time dep in the next image pipeline|2026-11-05
-pnpm|GHSA-w3rx-r6r6-pgpr|image-size 2.0.2; co-resolves with GHSA-5p2g, same absence of a patched release; build-time dep in the next image pipeline|2026-11-05
-python|CVE-2026-71554|h2 4.3.0 HTTP/2 advisory; transitive (httpx/httpcore + aiohttp stack); tracked for the next lock refresh|2026-11-05
+# ── 2026-09-23 re-triage — the list is now empty ─────────────────────────────
+# All 42 entries that stood here were re-verified against OSV and against the
+# versions the repo actually resolves, rather than against the claims in their
+# own reason strings. Every one of them was resolvable, so none was carried
+# forward. Nothing is suppressed today.
+#
+# The entries were re-checked ahead of the 2026-10-05 ecdsa expiry because
+# `load_ignores` fails closed: a lapsed entry aborts `validate-ignores` before
+# any ecosystem is scanned, which reds every open PR and leaves `main`
+# stale-green. That has happened here before (twelve expired suppressions,
+# #650), so the list is re-read on a schedule rather than on a failure.
+#
+# Recording what was measured, because most of these reason strings had
+# quietly stopped being true and re-reading them would have produced another
+# round of date bumps rather than fixes:
+#
+# • starlette (9 entries) claimed "slack-bot and teams-bot still cap fastapi
+# <0.137, which resolves fastapi 0.136.1 / starlette 1.0.0". Neither cap
+# exists: both pyproject files read fastapi >=0.111,<0.142 and both locks
+# resolve fastapi 0.141.1 with starlette 1.6.0. services/agents resolves
+# starlette 1.3.1. The highest fix among the nine is 1.3.1, so every
+# service was already clean.
+#
+# • idna CVE-2026-45409 claimed "only services/slack-bot still resolves
+# 3.13". slack-bot resolves 3.20, teams-bot 3.15, agents 3.18; the fix is
+# 3.15. (The CVE id is also not a resolvable OSV record — OSV carries this
+# as GHSA-65pc-fj4g-8rjx / PYSEC-2026-215 with the CVE as an alias.)
+#
+# • langchain / langgraph (9 entries) claimed pins of langchain <0.4.0,
+# langchain-openai <0.4.0 and langgraph <1.1.0 blocking a "0.3 to 1.x
+# migration". That migration has already happened: the pins read
+# langchain >=1.3.9, langchain-core >=1.2.22, langgraph >=1.0.10,
+# langgraph-checkpoint >=4.1.1, langgraph-sdk >=0.3.15, and the lock
+# resolves 1.4.2 / 1.6.4 / 1.2.12 / 4.2.0 / 0.4.5 respectively.
+# langchain-text-splitters (PYSEC-2026-77) is not in the tree at all.
+#
+# • langchain-core (4 entries) claimed the same blocked major bump. The lock
+# resolves 1.6.4; the fixes are 1.2.11 and 1.2.22.
+#
+# • weasyprint (3) and anyio (5) claimed they were "pinned by the committed
+# services/agents lock, needs a lock refresh". That lock already carries
+# weasyprint 70.0 and anyio 4.15.1, which are the fixed versions.
+#
+# • pydantic-settings claimed "slack-bot/teams-bot still resolve 2.14.1".
+# Both resolve 2.15.0; the fix is 2.14.2.
+#
+# • cryptography (3 entries) claimed the fix was "held by the <50 and <49
+# version caps". No such cap exists anywhere — every pin reads <51 — and
+# services/agents resolves 50.0.1, which is the fixed line.
+#
+# • aiohttp (3 entries) claimed 3.14.1 awaiting a lock refresh; the locks
+# resolve 3.14.3, which is the fix. h2 claimed 4.3.0; agents resolves
+# 4.4.1, which is the fix.
+#
+# • image-size (2 pnpm entries) claimed the advisory "reports no patched
+# release (patched_versions "<0.0.0"), so an override has no clean
+# target". GHSA-5p2g-fcmc-qvqq and GHSA-w3rx-r6r6-pgpr both record a
+# vulnerable range of "<= 2.0.2", and npm has published 2.0.3 and 2.0.4;
+# OSV reports both as unaffected. `first_patched_version` being null in
+# the GitHub advisory is not the same claim as no fix existing, and
+# reading it as such kept two high-severity advisories suppressed for a
+# fix that was already on the registry. A pnpm override now pins
+# >=2.0.4 <3, which @docusaurus/mdx-loader's ^2.0.2 range accepts.
+#
+# • ecdsa PYSEC-2026-1325 (CVE-2024-23342, Minerva timing attack on P-256)
+# was the one entry whose "no patched version available" was true: OSV
+# records the affected range as introduced at 0 with no fixed event,
+# because upstream states python-ecdsa offers no side-channel resistance
+# and will not fix it. It was also true that ecdsa was unconditional under
+# python-jose. So the dependency was removed instead of re-suppressed:
+# services/api now uses PyJWT, which it already depended on for the OIDC
+# and SAML paths, and python-jose, ecdsa and rsa all leave the tree.
+#
+# Verification for the empty list: all 20 Python targets were exported exactly
+# as `run_pip_audit` does and every resolved pin checked against OSV — zero
+# advisories, zero unscanned. `pnpm audit` reports zero advisories with the
+# image-size override (two high without it). `govulncheck` reports zero
+# findings across all six Go modules.
+#
+# Add entries back only with a justification that names the version actually
+# resolved and the fix that is genuinely out of reach. An expiry is a
+# re-review date, not a snooze.
diff --git a/scripts/tests/test_security_audit.py b/scripts/tests/test_security_audit.py
index 5fe1930f3..a1e57a9e8 100644
--- a/scripts/tests/test_security_audit.py
+++ b/scripts/tests/test_security_audit.py
@@ -12,6 +12,7 @@ import pytest
import sys
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+import security_audit
from security_audit import (
Finding,
Ignore,
@@ -24,6 +25,8 @@ from security_audit import (
load_ignores,
parse_govulncheck_json,
parse_pip_audit_json,
+ run_govulncheck,
+ run_pnpm_audit,
)
@@ -413,6 +416,32 @@ class TestExitCodeFor:
)
assert exit_code_for(r) == 1
+ def test_unscanned_alone_fails(self):
+ """A clean scan and a skipped scan must not both exit 0.
+
+ This is the property that makes an exit of 0 mean anything: without
+ it, a service the gate could not read is indistinguishable from a
+ service the gate read and cleared. `services/slack-bot` sat in that
+ state behind a stale poetry.lock while its lock resolved advisories
+ every other service had already moved past (#650).
+ """
+ r = Report(unscanned=["services/slack-bot: poetry export failed — NOT scanned"])
+ assert exit_code_for(r) == 1
+
+ def test_unscanned_fails_even_with_only_low_findings(self):
+ r = Report(
+ findings=[Finding("python", "low", "X", "p", "l", "t")],
+ unscanned=["services/api: pip-audit exited 2 — NOT scanned"],
+ )
+ assert exit_code_for(r) == 1
+
+ def test_warnings_alone_do_not_fail(self):
+ """Warnings are observations about a scan that happened; they must stay
+ non-fatal so that a coverage gap remains the only reason a finding-free
+ run can still exit 1."""
+ r = Report(warnings=["some non-fatal note"])
+ assert exit_code_for(r) == 0
+
# ─── Report property tests ───────────────────────────────────────────────────
@@ -446,3 +475,121 @@ class TestReportProperties:
]
)
assert len(r.low_info) == 3
+
+
+# ─── Coverage-gap tests: a scan that did not happen must not read as clean ───
+
+
+class _Proc:
+ """Minimal stand-in for subprocess.CompletedProcess."""
+
+ def __init__(self, returncode: int = 0, stdout: str = "", stderr: str = "") -> None:
+ self.returncode = returncode
+ self.stdout = stdout
+ self.stderr = stderr
+
+
+class TestPnpmCoverageGaps:
+ def test_unparseable_output_is_a_coverage_gap_not_a_warning(self, monkeypatch, tmp_path: Path):
+ """pnpm audit that returns garbage has audited nothing.
+
+ Recorded as `unscanned` so `exit_code_for` fails. As a `warning` the
+ arm exited 0 and printed "pnpm: 0 findings" for a workspace it had
+ never successfully read.
+ """
+ monkeypatch.setattr(
+ security_audit.subprocess,
+ "run",
+ lambda *a, **k: _Proc(returncode=1, stdout="<html>not json</html>"),
+ )
+
+ report = run_pnpm_audit(tmp_path, [])
+
+ assert report.unscanned, "unparseable pnpm output must be recorded as a coverage gap"
+ assert not report.findings
+ assert exit_code_for(report) == 1
+
+ def test_successful_audit_records_no_gap(self, monkeypatch, tmp_path: Path):
+ monkeypatch.setattr(
+ security_audit.subprocess,
+ "run",
+ lambda *a, **k: _Proc(returncode=0, stdout='{"advisories": {}}'),
+ )
+
+ report = run_pnpm_audit(tmp_path, [])
+
+ assert report.unscanned == []
+ assert exit_code_for(report) == 0
+
+
+class TestGoCoverageGaps:
+ @staticmethod
+ def _module(tmp_path: Path) -> Path:
+ mod = tmp_path / "services" / "ingest"
+ mod.mkdir(parents=True)
+ (mod / "go.mod").write_text("module example.com/ingest\n")
+ return tmp_path
+
+ def test_unexpected_exit_is_a_coverage_gap(self, monkeypatch, tmp_path: Path):
+ """govulncheck exits 0 (clean) or 3 (vulnerabilities found).
+
+ Any other code means the module never got analysed — a build failure
+ or a missing toolchain. Treated as a coverage gap for the same reason
+ as the Python arm: otherwise the arm prints "go: 0 findings" for a
+ module it could not read.
+ """
+ root = self._module(tmp_path)
+ monkeypatch.setattr(
+ security_audit.subprocess,
+ "run",
+ lambda *a, **k: _Proc(returncode=1, stderr="build failed: no required module provides package"),
+ )
+
+ report = run_govulncheck(root, [])
+
+ assert report.unscanned, "a failed govulncheck run must be recorded as a coverage gap"
+ assert "services/ingest" in report.unscanned[0]
+ assert exit_code_for(report) == 1
+
+ def test_missing_binary_is_a_coverage_gap(self, monkeypatch, tmp_path: Path):
+ root = self._module(tmp_path)
+
+ def _raise(*a, **k):
+ raise FileNotFoundError("govulncheck")
+
+ monkeypatch.setattr(security_audit.subprocess, "run", _raise)
+
+ report = run_govulncheck(root, [])
+
+ assert report.unscanned
+ assert exit_code_for(report) == 1
+
+ def test_clean_module_records_no_gap(self, monkeypatch, tmp_path: Path):
+ root = self._module(tmp_path)
+ monkeypatch.setattr(security_audit.subprocess, "run", lambda *a, **k: _Proc(returncode=0, stdout=""))
+
+ report = run_govulncheck(root, [])
+
+ assert report.unscanned == []
+ assert report.findings == []
+ assert exit_code_for(report) == 0
+
+ def test_findings_exit_code_3_still_counts_as_scanned(self, monkeypatch, tmp_path: Path):
+ root = self._module(tmp_path)
+ stdout = "\n".join(
+ [
+ '{"osv": {"id": "GO-2026-0001", "aliases": ["CVE-2026-0001"], "summary": "boom"}}',
+ '{"finding": {"osv": "GO-2026-0001"}}',
+ ]
+ )
+ monkeypatch.setattr(
+ security_audit.subprocess,
+ "run",
+ lambda *a, **k: _Proc(returncode=3, stdout=stdout),
+ )
+
+ report = run_govulncheck(root, [])
+
+ assert report.unscanned == []
+ assert [f.vuln_id for f in report.findings] == ["GO-2026-0001"]
+ assert exit_code_for(report) == 1
diff --git a/services/api/Dockerfile b/services/api/Dockerfile
index 0205c1232..d9c8d7f71 100644
--- a/services/api/Dockerfile
+++ b/services/api/Dockerfile
@@ -43,7 +43,6 @@ RUN set -eux; \
"alembic>=1.13,<2" \
"redis[hiredis]>=5.0.4,<8.0.0" \
"httpx>=0.27,<0.29" \
- "python-jose[cryptography]>=3.3,<4" \
"PyJWT>=2.8,<3" \
"cryptography>=42,<43" \
"bcrypt>=4.1,<5" \
diff --git a/services/api/app/api/v1/deps.py b/services/api/app/api/v1/deps.py
index f9aca01b0..b476ba8fd 100644
--- a/services/api/app/api/v1/deps.py
+++ b/services/api/app/api/v1/deps.py
@@ -27,7 +27,7 @@ from typing import Annotated
from fastapi import Depends, HTTPException, Security, status
from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
-from jose import JWTError
+from jwt import PyJWTError
from sqlalchemy import select, update
from sqlalchemy.ext.asyncio import AsyncSession
@@ -218,7 +218,7 @@ async def get_current_user(
token_type: str = payload.get("type", "access")
if user_id is None or token_type != "access":
raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail="Invalid token")
- except JWTError as e:
+ except PyJWTError as e:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Could not validate credentials",
diff --git a/services/api/app/api/v1/endpoints/graph_ws.py b/services/api/app/api/v1/endpoints/graph_ws.py
index 29fb68279..29a8dc07b 100644
--- a/services/api/app/api/v1/endpoints/graph_ws.py
+++ b/services/api/app/api/v1/endpoints/graph_ws.py
@@ -59,7 +59,7 @@ from fastapi import (
WebSocketDisconnect,
status,
)
-from jose import JWTError
+from jwt import PyJWTError
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
@@ -139,7 +139,7 @@ async def _authenticate_ws(
# JWT path
try:
payload = decode_token(token)
- except JWTError:
+ except PyJWTError:
await websocket.close(code=status.WS_1008_POLICY_VIOLATION, reason="invalid token")
return None
sub = payload.get("sub")
diff --git a/services/api/app/core/security.py b/services/api/app/core/security.py
index aa863fe88..d26ba33b0 100644
--- a/services/api/app/core/security.py
+++ b/services/api/app/core/security.py
@@ -8,7 +8,7 @@ from datetime import UTC, datetime, timedelta
from typing import Any
import bcrypt
-from jose import jwt
+import jwt
from app.core.config import settings
diff --git a/services/api/pyproject.toml b/services/api/pyproject.toml
index 6e9f5c4f3..bacf9b9e5 100644
--- a/services/api/pyproject.toml
+++ b/services/api/pyproject.toml
@@ -20,8 +20,13 @@ asyncpg = ">=0.29.0,<0.32"
alembic = "^1.13.1"
redis = { version = ">=5.0.4,<9.0.0", extras = ["hiredis"] }
httpx = ">=0.27,<0.29"
-python-jose = { version = "^3.3.0", extras = ["cryptography"] }
-PyJWT = "^2.8.0" # Used by OIDC and SAML auth flows (separate from python-jose)
+# PyJWT is the only JWT implementation in this service: first-party access and
+# refresh tokens, realtime tickets, and the OIDC / SAML assertion paths. It
+# replaced python-jose, whose unconditional `ecdsa` requirement carried
+# CVE-2024-23342 (Minerva timing attack on P-256) with no patched release
+# upstream — the maintainers state python-ecdsa offers no side-channel
+# resistance and will not fix it. Nothing here needed jose's JWE/JWS surface.
+PyJWT = "^2.8.0"
cryptography = ">=42,<51" # Direct dep for credential vault Fernet
bcrypt = ">=4.1.2,<6.0.0" # Used directly by app.core.security (passlib is no longer compatible with bcrypt >= 4)
python-multipart = ">=0.0.9,<0.0.33"
diff --git a/services/api/tests/api/v1/endpoints/test_realtime_ticket.py b/services/api/tests/api/v1/endpoints/test_realtime_ticket.py
index 4cd074320..321aba408 100644
--- a/services/api/tests/api/v1/endpoints/test_realtime_ticket.py
+++ b/services/api/tests/api/v1/endpoints/test_realtime_ticket.py
@@ -27,13 +27,13 @@ from __future__ import annotations
import uuid
+import jwt
import pytest
from app.api.v1.deps import CurrentUser
from app.api.v1.endpoints import realtime as realtime_ep
from app.core.config import DEV_REALTIME_TICKET_SECRET, Settings, realtime_ticket_secret, warn_if_insecure_defaults
from app.core.security import REALTIME_TICKET_AUDIENCE
from fastapi import HTTPException
-from jose import jwt
# A real 64-char secret used across the "happy path" tests — anything outside
# INSECURE_SECRET_KEY_DEFAULTS and non-empty resolves as configured.
@@ -157,7 +157,7 @@ async def test_mint_rejects_foreign_audience(configured):
"""A ticket scoped to the realtime edge must not validate for any other
audience — guards against a stolen ticket being replayed at the API."""
ticket = await realtime_ep.mint_realtime_ticket(_user())
- with pytest.raises(jwt.JWTError):
+ with pytest.raises(jwt.PyJWTError):
jwt.decode(
ticket.token,
_REAL_SECRET,