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 — aws-credential-exfil
AWS credential exfiltration. Long-lived IAM access keys belonging to a CI service account (
ci-svc-prod) are used from an ASN the workload has never previously contacted, immediately followed bys3:ListBuckets+ bulks3:GetObjectagainst a tagged-confidential bucket.
- Fixture:
examples/alerts/aws-credential-exfil.json - MITRE techniques: T1552, T1567, T1078.004
- Severity at intake:
critical - Confidence band (sandbox): 100/100,
high
Run it
# Offline simulator:
aisoc-sandbox demo --scenario aws-credential-exfil
# Real stack (after `make up`, with `pip install -e packages/aisoc-cli`):
aisoc submit examples/alerts/aws-credential-exfil.json
What the agent does, step by step
Step 0 · DetectAgent · detect
Three CloudTrail events ingest as a single fused incident — the AWS
GuardDuty + native AiSOC detections both fire on the
UnauthorizedAccess:IAMUser/AnomalousASN pattern. Fusion deduplicates
the three events and lifts the alert to severity=critical because
the bucket carries the Classification=Confidential tag.
Step 1 · TriageAgent · triage
The graph walk identifies that ci-svc-prod is a long-lived service
account whose key has never rotated, and that the access pattern
(sts:GetCallerIdentity → s3:ListBuckets → bulk s3:GetObject)
matches three public access-key-leak postmortems from Q1 2026.
Decision: Confidence high (100/100).
Step 2 · HuntAgent · hunt
Sweeps the last 24 h for any other API calls from 198.51.100.77 or
using AKIAEXAMPLE7QWERTY. The hunt confirms the key has been used
from exactly one new IP and against exactly one new bucket — clean
exfil signature, no broader compromise yet.
Step 3 · RespondAgent · respond
Proposes a five-action containment burst — every action sits inside
the human_approval_required gate because severity is critical:
iam.access_key.deactivate({"access_key_id": "AKIAEXAMPLE7QWERTY"})iam.user.attach_deny_policy({"user": "ci-svc-prod"})s3.bucket.block_public_access({"bucket": "acme-prod-customer-pii"})case.create(...)at severitycriticalcase.notify({"channel": "pagerduty", "service": "soc-tier1"})
What the analyst would do next
- Approve
iam.access_key.deactivatefrom the responder PWA. The deny policy can wait until the on-call confirms there's no automated rotation in flight. - Look at the bucket's
s3:GetObjectaudit log for the full set of objects accessed — the alert shows one specific object, but theadditionalEventData.totalObjectsListedInWindow=4137suggests the actor was iterating. - Trigger the secrets-rotation playbook for every IAM key issued in the last 90 days.
- Open a separate case for
ci-svc-proditself — long-lived keys on a CI account is the root cause, and the org should migrate to short-lived OIDC tokens.