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 pack
For a reviewer doing diligence on AiSOC, in the order a security questionnaire usually walks them.
| Document | Answers |
|---|---|
| Architecture and data flow | What the components are, which trust boundaries they sit across, and where every class of data goes. |
| Data handling by deployment mode | What leaves your perimeter under each of the four ways AiSOC can be run, and what is retained where. |
| Platform threat model | STRIDE over the platform, with the credential vault as the top asset. |
| Agent and tool threat model | STRIDE over the prompt and tool boundaries, both of which take attacker-influenced input. |
| Access governance | Attribute conditions, time-boxed elevation and per-service workload identities, and where each is narrower than its heading sounds. |
| Connector least privilege | The minimum vendor scope each connector needs. |
| Security questionnaire | Standard control questions, answered, with the gap list stated rather than omitted. |
| Sub-processors | What this repository can and cannot establish about the hosted service. |
Leaving access-governance.md out of this table is how the questionnaire
came to answer "no reader" about three controls that had one: a reviewer
walking the pack never reached the page that documents them.
The one rule these documents follow
Every statement names the file or the CI job that makes it true, and where a control is narrower than its heading sounds, the limit is in the same paragraph.
That rule exists because this project has broken it. Four security documents
once described controls that did not exist in the tree: AES-256-GCM backup
encryption where the script only gzipped and uploaded, envelope encryption
cited as the mitigation for a database dump while EnvelopeCipher had zero
callers, one unbroken distributed trace with neither end of the Kafka spine
instrumented, and per-tenant retention marked GATED by a test that asserted
only that the purge SQL parses. Each was corrected, and then implemented.
So a claim here without a link beside it is a bug in this pack, and reporting one is welcome.
Verifying any of this yourself
Nothing here asks to be taken on trust. The repository is the evidence:
git clone https://github.com/beenuar/AiSOC
cd AiSOC
# Which CI job proves which product claim — 293 rows, every one naming a gate.
python3 scripts/check_claim_gate_matrix.py
# The authorization surface, counted rather than described.
python3 scripts/check_route_auth.py # every route authenticates, or is listed with a reason
python3 scripts/check_route_authz.py # every state-changing route authorizes, or is listed
python3 scripts/check_one_permission_model.py --check
# Nothing is pointed at the internet by default.
python3 scripts/check_default_egress.py
# Every gate named in this pack is actually wired to a workflow.
python3 scripts/check_gate_coverage.py
docs/audit/CLAIM_TO_GATE_MATRIX.md is the
long form: one row per published claim, each naming the job that fails when the
claim stops being true. Rows marked PARTIAL name their own gap.
What this pack is not
- Not an audit report. No third-party audit has been commissioned against
AiSOC. The reasoning, the trigger condition and the sequencing are recorded
in ADR-0002.
That ADR also asserts a CI check that does not exist — see
the questionnaire, which states
the real position. Surfacing that is this pack's own rule working: the
sentence was written from the ADR and caught by
scripts/check_security_pack_links.pybefore it shipped. - Not a penetration-test report. None has been commissioned. What exists instead is a public advisory history: 16 advisories published at the time of writing, covering findings from both external reports and internal audits, each with the fix and the regression guard. See the questionnaire and the advisory list.
- Not a statement about a deployment you have not seen. Most of these controls are properties of the software. Whether your operator configured them is a question about your deployment, and the data-handling page says which is which.