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.
Python agentic-aiai-securityalert-triageblue-teamcybersecuritydetection-engineeringincident-responselanggraphllm-agentsmcpmitre-attackmodel-context-protocolpurple-teamsecurity-operationsself-hostedsiemsoarsocthreat-detectionthreat-intelligence
AiSOC — known gaps & follow-ups
State as of 2026-10-03, after the fix session (see fix-log-2026-10.md).
Data / ingestion
- ClickHouse lake empty (
aisoc.raw_events= 0 rows). Lake query chain is fixed end-to-end (available: true), but nothing has indexed events into the lake yet. Until ingestion backfills, agent investigations run with thin evidence and the reports will be model-generated rather than evidence-grounded. Next step (needs a decision): verifyaisoc-ingest→ kafka → lake sink path and trigger a backfill window. aisoc-ingestreportssubscriptions: {}at /readyz — consumer detached or not configured on this build. Confirm intended consumption topology before forcing a restart.- OpenSearch vs ClickHouse ambiguity. Both defined; API lake routes
point at ClickHouse (
CLICKHOUSE_HOST: clickhouse). OpenSearch runs but holds no event indices. Decide whether it stays (retire it if unused).
Upstream sync obligation
All fixes are on the fork working tree. The upstream repo
(beenuar/AiSOC) must receive these via PR or they regress on the next
image pull. Priority upstream candidates:
- proxy auth fixes (cases/playbooks/graphql → agents)
_normalize_status+ forward-monotonic ladder- investigations list fallback to the postgres ledger
- React #31 responder-card coercion + overview timeline fetch
- ledger
new→triagedstatus lift + approvals principal permission fix - alerts assignee alias + model_fields_set patch
GPU / model
- T4 16 GB:
qwen3:8bthinking chains are unusable (60 s+ per call). Current:qwen2.5:7b-instruct(sub-second to ~6 s). Revisit only with a bigger GPU or litellm timeout raise + streaming. ollama n_slots=1: single request at a time; concurrent investigations queue. Fine for this footprint, known ceiling.
UI follow-ups (not blocking)
- Report tab auto-refresh after a live run completes (currently fetched once on completion; a late report.md write isn't picked up).
- Audit log entries render
kind/agent/summarystrings only — any future object-valued summary field must be coerced (React #31 class bug).
Deliberately NOT fixed (by design — do not "fix" again)
- Backwards case status transitions → 422 (ladder guard).
escalateverdict leaving alertstatus='new'(human queue stays open).- Source-HIDS SIEM writeback: no disposition-write arm exists for the HIDS connector; the execute flag changing nothing is correct.
- Approvals 502 when approver truly lacks
actions:execute:high— the gate working; the fixed bug was only the empty-permission forwarding.