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.
Walkthrough — github-token-theft
GitHub PAT leaked. A personal-access token tied to
[email protected]is used from an ASN the org has never previously contacted, against six private repositories in 11 seconds. The token'sreposcope is wide enough to ex-filtrate every source tree Carol can read.
- Fixture:
examples/alerts/github-token-theft.json - MITRE techniques: T1078, T1555, T1567
- Severity at intake:
high - Confidence band (sandbox): 87/100,
high
Run it
# Offline simulator:
aisoc-sandbox demo --scenario github-token-theft
# Real stack (after `make up`, with `pip install -e packages/aisoc-cli`):
aisoc submit examples/alerts/github-token-theft.json
What the agent does, step by step
Step 0 · DetectAgent · detect
The github-pat-bulk-clone detection fires on six private-repo clones
from the same hashed_token_id within an 11-second window — well
inside the per-token velocity threshold. Fusion correlates against the
threat-intel feed and notes that the source ASN has appeared in two
other OSS supply-chain incidents in the last 90 days.
Step 1 · TriageAgent · triage
The graph walk pivots from carol to her org memberships and the
repos she has push access to. The triage rationale notes that
infostealer behaviour against the OS-native keystore is consistent
with the access pattern (machine compromised, PAT lifted, used
immediately).
Decision: Confidence high (87/100).
Step 2 · HuntAgent · hunt
Sweeps for any other action by Carol's PAT or session in the last
24 h — including pushes, branch creations, and Action triggers. The
hunt confirms the actor has only cloned so far; no malicious code has
been pushed yet, which means there's a window to revoke before
secrets are siphoned out of acme/secrets-vault.
Step 3 · RespondAgent · respond
Proposes a five-action containment that closes both the credential path and the secrets-at-rest path:
github.pat.revoke({"user": "carol", "token_prefix": "ghp_xxxAAAAA"})github.user.session.revoke_all({"user": "carol"})github.org.audit_member({"user": "carol", "lookback_days": 30})secrets.rotate({"repositories": ["acme/secrets-vault", "acme/iam"]})case.create(...)at severityhigh
Decision: 5 actions proposed; awaiting analyst approval (severity high).
What the analyst would do next
- Approve
github.pat.revokefirst — it has zero blast radius for anyone other than Carol. - Look at every commit + Action run + secret access from the cloned
repos in the last 7 days; even read-only cloning gives the actor
any secret that was committed historically (which is why
secrets.rotateis part of the response, not optional). - Reach Carol via an out-of-band channel (phone, in-person) — not Slack or email — to confirm she didn't run the clones and to investigate her machine.
- Open a separate workstream on the org's PAT lifecycle — the org
should be on fine-grained GitHub App tokens, not personal-access
tokens with
reposcope.