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.
Security batch progress
Thirteen defects found by reading the code at v14.0.0 (commit 079fdb6). None
had been exercised against a running deployment, so every item starts with a
reproduction that fails on the untouched tree for the stated reason.
Two of the thirteen are also items in
plans/aisoc_parity_plan.plan.md. They are
done once, here, and the parity tracker points at this file rather than
repeating the work.
D0. This batch is public, by the maintainer's decision
The source brief specified coordinated disclosure: a local branch pushed
nowhere, a tracker under .private/, and advisory drafts handed over in chat
for the maintainer to publish before any code became visible.
The maintainer chose the opposite: ordinary public pull requests, merged and
released as they land, with advisories filed afterwards against the release.
The reasoning is that this is a self-hosted product whose defaults are readable
by anyone who opens .env.example, so the fix reaching self-hosters sooner is
worth more than the delay buying secrecy that the repository does not really
have.
What that changes, and what it does not:
- No
.private/directory. This file is the tracker and it is committed. - Commit messages stay neutral and carry no exploit detail, payloads or advisory wording. That costs nothing, keeps the history readable, and leaves the option of filing advisories against these commits later.
- Advisory drafts are still produced for every defect: reproduction, root cause, affected versions, a suggested CVSS v3.1 vector, and whether the fix is breaking.
- S1 and S13 both break callers, so this batch lands as a major.
Pre-existing state, captured before any code was written
Measured on origin/main at 079fdb69 with the working tree stashed, so the
numbers describe the untouched tree.
Gates: 77 of 84 pass
The seven that do not, and why none is a defect in the tree:
| Gate | Exit | Reason |
|---|---|---|
check_attribution.py |
2 | Refuses to scan without --all, --diff-base, --commits or --text-file. |
check_container_egress.py |
2 | Refuses without --service: "a probe that covers nothing must not report clean". |
check_published_secrets.py |
2 | Refuses without --check-env, --check or --list-guarded. |
check_codeql_alerts.py |
2 | Needs GITHUB_TOKEN with security-events: read. Refuses rather than reporting clean about alerts it never read. |
check_main_run_cancellations.py |
2 | Needs a token for the Actions API, same refusal. |
check_required_check_substance.py |
2 | Needs a token for the Actions API, same refusal. |
check_mypy_baseline.py |
1 | Environmental, see D14. |
Six of the seven are the gate contract working: they decline to render a
verdict about a corpus they could not read. Only check_mypy_baseline reports
a finding, and it does not reproduce in CI.
Console suite
apps/web: 79 files, 816 tests, zero failures. npx tsc --noEmit clean.
npx eslint . reports 0 errors and 76 warnings, all pre-existing.
Deviations: where the brief and the tree disagree
D1. S1c is 61 call sites, not 27
The brief names RBACView.tsx and "any other header-less call you find".
Counting by hand with a grep for fetch('/api/v1 gives 27. The gate written
for the item finds 61 across 21 files, and the extra 34 are not noise:
- template-literal URLs behind a base variable, which the grep pattern cannot
see:
fetch(\${API}/api/v1/honeytokens/${id}`)` on the honeytokens and purple-team pages (9 calls); - six calls inside
apps/web/src/lib/api.tsitself, five of which sentX-Tenant-Idand no token, which reads like an identity and is not one; - sixteen
useSWRsites whose fetcher is a single-argument function and therefore anonymous by construction.
D2. require_permission_db has 27 call sites, not 31
All 27 are in-body await <principal>.require_permission_db(...); there are
zero as a Depends. They live in four modules only: autonomy_policy.py
(9), tenant_skills.py (9), business_context.py (5), mcp_servers.py (4).
The static figure in the brief is exact: 272, being 256 Depends plus 16
direct calls.
D3. S3's "no authorization decision" is narrower than written
All four routes in services/api/app/api/v1/endpoints/shifts.py do take
current_user: AuthUser, so they authenticate. What is absent is any
require_permission(...) and any tenant scoping, so the finding is
cross-tenant read and write of one shared in-process list by any authenticated
principal. The effect stands; the mechanism in the brief does not.
D4. S7's audit-read restriction is not a wildcard check
services/api/app/api/v1/endpoints/audit.py contains no "*" comparison.
Both read routes use Depends(require_permission("audit_log:read")). The
outcome the brief describes holds and is stronger than stated: audit_log:read
appears in no role's permission list in ROLE_PERMISSIONS, so only the two
roles holding ["*"] can read the log. A tenant_admin gets 403 because the
permission is unreachable, not because a wildcard was demanded.
D5. The production compose file already exists and nothing selects it
S1's fix is not "write a production compose". docker-compose.prod.yml sets
ENVIRONMENT: production and AISOC_DEV_MODE: 0 as literals on all six Python
services. But Makefile:102 (up:) and Makefile:142 (up-full:) never pass
-f docker-compose.prod.yml, and install.sh and install.ps1 contain zero
occurrences of either variable. The file is correct and unreachable from every
documented path.
D6. realtime's push routes are worse than written
The brief says the tenant comes "from a header or query parameter with no
credential". services/realtime/src/push.ts:92 reads x-tenant-id, falls back
to the tenant_id query parameter, and then falls back to the literal string
'default'. The three POST /v1/push/* routes carry only a rate limiter.
D7. Ten services default AISOC_DEV_MODE=1, and api is not one of them
docker-compose.yml sets AISOC_DEV_MODE: ${AISOC_DEV_MODE:-1} on fusion,
agents, osquery-tls, actions, connectors, threatintel, ueba, honeytokens,
purple-team and slack-bot. The API service has no AISOC_DEV_MODE at all; it
reaches the same bypass through ENVIRONMENT: ${ENVIRONMENT:-development} and
AUTH_BYPASS_ENVIRONMENTS.
D14. check_mypy_baseline fails locally and passes in CI
Locally it reports (unmanaged)/scripts/run_evals.py: 15 index finding(s), baseline records 13. CI is green at the same commit (28 success, 2 skipped on
079fdb69), and the local mypy is 2.3.1, which is what ci.yml:234 pins.
The gate deliberately removes two sources of variance with
--no-site-packages --no-incremental, documented at ci.yml:272. It does not
pin the interpreter, and this machine runs Python 3.12.6 against CI's 3.11.
Recorded as environmental rather than fixed, and worth noting as a gap in a
gate whose whole purpose is reproducibility.
S1. Anonymous admin in the documented deployment
Reproduced and confirmed live on origin/main. The crux is byte-identical, not
merely similar: dev_auth.py:29 DEMO_TENANT_ID and bootstrap_admin.py:59
DEFAULT_TENANT_ID are both 00000000-0000-0000-0000-000000000001, so the
anonymous principal is an admin in the same tenant the real operator is
bootstrapped into.
- S1c Console credential on every call (#1069)
- S1a The bypass needs an explicit opt-in, refuses a reachable bind address, and the demo identity has its own tenant (#1070)
- S1b The nine vendored shims, the three
require_service_authcopies, slack-bot and realtime push (#1070) - S1d Production-class defaults on every documented path (#1073)
S2 to S13
- S2 Copilot conversations and saved hunt searches are tenant-scoped
- S3 Fabricated shared shift state
- S4 STIX and TAXII demo data served outside demo mode
- S5 Login throttling and lockout
- S6 SSO stub identities deleted
- S7 Audit blind spots
- S8 Tenant isolation beyond query predicates
- S9 State-changing routes with no authorization decision
- S10 Attacker-controlled alert text can close and then silence alerts
- S11 Plaintext service traffic
- S12 Default Caldera API key
- S13 A custom role cannot restrict a user below their built-in role
Maintainer actions
- Check whether the managed instance runs with a dev-class
ENVIRONMENTor serves more than one real tenant. If it does, assess exposure for S1 to S4 and rotate anything those routes could reach. (Named neutrally on purpose:scripts/check_hosted_hostname.pykeeps the managed hostname out of this repository, and a tracker is not a reason to exempt it.) - Reserve the PyPI and npm names the packages will use (
aisocon PyPI belongs to an unrelated project). - File the advisories against the release once the batch has landed.
D9. Flipping the default to production stops the API booting, until two more secrets exist
Found by starting the stack, not by a test. enforce_secure_defaults refuses
to boot when METRICS_TOKEN or JWT_SECRET is empty and the environment is
not development-class, and ensure_env.py generated neither, because in a
development-class environment that check only warns. So the posture change on
its own would have turned "anonymous administrator" into "the API will not
start". Both are generated now, taking make env from twelve secrets to
fourteen.
Two adjacent defects surfaced in the same hour, both of the same shape as one
this repository has hit before — a value in .env that cannot reach the
container that needs it:
- neither secret was declared on the
apiservice indocker-compose.yml, so generating them was not sufficient; ingest-workerhardcodedJWT_SECRET: dev_secret_key_change_in_productionas a bare literal. That value is inINSECURE_SECRET_KEY_DEFAULTS, so anyone who read this repository held the key the service verified with, and no.envvalue could displace it.
D10. The first live verification graded a cached image
The probe stack came up from ghcr.io/beenuar/aisoc-core-api:latest, which
was already on the machine, so the first three posture results described the
pre-fix API rather than this branch. They looked plausible: production refused
and development admitted, which is true of both trees. The tell was the third
case, where a published bind address failed to refuse — a control that does
not exist in the old code. Rebuild before probing, and prefer a probe that has
a case the old code cannot pass.
D8. Two documents publish the claim-matrix row count in prose
ROADMAP.md and apps/docs/docs/compliance/evidence-pack.md both state the
figure, and scripts/check_readme_figures.py compares them against the live
table. So every PR that adds a claim row moves three files, not one. The gate
caught it on the first attempt here, which is what it is for.
Notes for the next session
The batch order is dependency-driven, not severity-driven. S1c had to land before S1 because closing the anonymous path turns 61 console calls into 401s, and the console had to be able to authenticate first.
S1 split into S1a/S1b (the code that decides whether to admit an uncredentialed caller, across all eleven services at once so the vendored copies and their cross-references move together) and S1d (the deployment paths). They are separable because S1a/S1b close the hole and S1d makes the documented experience match; shipping the first without the second leaves a correct default and a stale guide, which is the safe order.
Two things found while fixing S1b that neither brief raised:
- the API's push proxy stamped
X-AiSOC-Internal-Tokenwhile realtime'srequireInternalreadsx-internal-token, so the token it sent was never checked by anything; userOfinservices/realtime/src/push.tspreferred a caller-supplied body field over the stampedX-User-Id, under a comment asserting that the API gateway validates it. It does not.
Complete — released as v15.0.0 (2026-10-01)
All thirteen defects are fixed and on main. Seventeen pull requests;
main green; claim-to-gate matrix at 255 rows, all GATED.
| Defect | PR |
|---|---|
| S1a/S1b anonymous bypass + demo tenant split | #1072 |
| S1c console credentials on all 61 calls | #1069 |
| S1d production posture on documented paths | #1074 |
| S2 copilot + saved hunts tenant-scoped | #1079 |
| S3/S4 fabricated shift + STIX data | #1077 |
| S5 login throttling and lockout | #1075 |
| S6 SSO stub identities deleted | #1076 |
| S7 audit attribution + tenant-admin reads | #1080 |
S8 RLS on the six unpoliced mssp_* tables |
#1084 |
| S9 route authorization decisions | #1081 |
| S10 human-confirmed priors only | #1082 |
| S11/S12 TLS everywhere + Caldera default | #1078 |
| S13 database-backed RBAC at every route | #1085 |
Release #1086 · advisories
#1087 → docs/security/v15-advisory-drafts.md.
Published and verified at the registry, not from a green workflow: the
release body survived at 52,623 characters (the cap is 125,000), and all
39 first-party image references resolve in GHCR with v15.0.0 manifests
returning HTTP 200.
What this batch actually taught
Not a category of bug. A shape: a control that exists, passes its tests,
and never runs. 275 route dependencies read a hardcoded map while the
console wrote to a database nothing consulted. A prompt-injection flag was
read from a field no code ever set. Thirteen Kafka clients each took the
library default. An ImportError branch nobody could avoid was the only
reachable path through the SAML assertion consumer.
Three mistakes of my own are worth more than the fixes:
- A real database caught what a static gate could not. The obvious
MSSP policies are mutually recursive; Postgres answers
infinite recursion detectedon the firstSELECTwhilecheck_rls_policy_shape.pypassed throughout. Shape is not liveness. - Adding one keyword broke four writer signatures and no test failed on
the exception, because the call sits inside
contextlib.suppress. What surfaced was a zero write-count in an unrelated replay test three files away. - The first design for database-backed permissions was wrong. Resolving inside the permission check would have made a transient database fault deny every request on the platform. It resolves at authentication and fails open to the static map, deliberately.
Still open
e2e.yml :: Buyer journey is intermittently red and not a required
check. next/font/google fetches from Google at compile time, so the
real dependency is the network. Removing the pnpm store cache produced one
green run and the next commit failed again on a different font, so that
hypothesis is disproved and recorded as such in the workflow. The fix is
NEXT_FONT_GOOGLE_MOCKED_RESPONSES or self-hosting the five faces through
next/font/local.