fix(security): close LOW ASVS findings — env bypass, SSE/deployment authz, CN fail-close, password byte-limit, exception leaks, BUG-12..16
Auth/session (V2.1.7, V4.1.5, V4.1.6, V2.1.4/V2.1.5): - env secret validation no longer bypassed by attacker-injectable PYTEST* env; gated on explicit DECNET_TESTING=1 (set only in conftest). - must_change_password now enforced on the SSE header-JWT path, not just ticket mint. - GET /system/deployment-mode requires viewer auth (was leaking role + topology size). - CreateUser/ResetUser passwords min_length=12; passwords >72 bytes rejected explicitly instead of bcrypt silently truncating. Swarm ingestion (V9.1.3, BUG-16): - Log listener hard-rejects peers with unparseable/empty cert CN (fail closed, ingests nothing) instead of tagging 'unknown'. - Shutdown handlers no longer swallow real errors (narrowed to CancelledError). Info leakage (V7.1.2, V14.1.2): - Exception text sanitized on swarm-update, health, tarpit, realism, file-drop, blank-topology endpoints (raw tc/docker stderr, DB/Docker errors logged server-side, generic detail returned). pyproject license corrected to AGPL-3.0. Correctness (BUG-12..16): - BUG-12 atomic credential upsert (UNIQUE constraint + IntegrityError retry, consistent principal_key canonicalization). - BUG-13 rule-tail watermark uses >= with seen-id dedup (no same-second drop). - BUG-14 worker wake cleared before wait (no lost wake during tick). - BUG-15 intel gather tolerates an unexpected provider raise. - BUG-16 see above. Already-closed (verified, no change): V2.1.6, V5.1.3, V9.1.2. Accept-risk + documented: V2.1.8 cache window, V3.1.3 idle timeout. Tests added for every fix; unanimous adversarial review after two refute-fix rounds.
This commit is contained in:
@@ -37,6 +37,17 @@ oauth2_scheme = OAuth2PasswordBearer(tokenUrl="/api/v1/auth/login")
|
||||
# Per-request user lookup was the hidden tax behind every authed endpoint —
|
||||
# SELECT users WHERE uuid=? ran once per call, serializing through aiosqlite.
|
||||
# 10s TTL is well below JWT expiry and we invalidate on all user writes.
|
||||
#
|
||||
# REVOCATION-WINDOW NOTE (V2.1.8, accept-risk): these per-process caches bound
|
||||
# how long a *stale* user/username/denylist row can be served. Single-process:
|
||||
# a role downgrade or password change is reflected within ~_USER_TTL (10s) even
|
||||
# if the in-process invalidate_user_cache hook is missed; the local write path
|
||||
# invalidates immediately. Multi-worker (gunicorn/uvicorn --workers>1) gives
|
||||
# each worker its own cache, so the worst-case cross-worker staleness is also
|
||||
# ~10s and would need a shared cache (Redis) to collapse to zero. This staleness
|
||||
# is NOT the authoritative revocation control: JWT bulk-revoke via the user's
|
||||
# tokens_valid_from cutoff (enforced in _resolve_token) is the hard cutoff and
|
||||
# is unaffected by these caches.
|
||||
_USER_TTL = 10.0
|
||||
_user_cache: dict[str, tuple[Optional[dict[str, Any]], float]] = {}
|
||||
_user_cache_lock: Optional[asyncio.Lock] = None
|
||||
@@ -389,6 +400,11 @@ def require_stream_role(*allowed_roles: str):
|
||||
header_token = _bearer_from_header(request)
|
||||
if header_token:
|
||||
_user_uuid, user = await _resolve_token(header_token)
|
||||
if user.get("must_change_password"):
|
||||
raise HTTPException(
|
||||
status_code=status.HTTP_403_FORBIDDEN,
|
||||
detail="Password change required before accessing this resource",
|
||||
)
|
||||
if user["role"] not in allowed_roles:
|
||||
raise HTTPException(
|
||||
status_code=status.HTTP_403_FORBIDDEN,
|
||||
|
||||
Reference in New Issue
Block a user