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
{
"schema": "aisoc.autonomy_performance/v1",
"status": "pending",
"reason": "No evaluation provider is funded in this repository, so no auto-closure precision has been measured for any model. The honest content of this file is zero rows. A fabricated row, or a row scoring 0.000, would read as a measured grade and would be the defect app/confidence/eligibility.py exists to prevent. Every configuration therefore resolves to UNMEASURED, which is additive by default and denied under AISOC_AUTONOMY_REQUIRE_MEASURED=1.",
"produced_by": "scripts/run_verdict_eval.py --out <scorecard.json>, promoted by hand after review",
"promotion_rules": [
"A row may only be added from a scorecard whose mode is 'live' and whose llm_calls_placed is greater than zero. A deterministic run grades the offline fallback and says nothing about a model.",
"auto_close_precision is the scorecard's auto_close_precision, computed over auto_close_considered. It is not headline accuracy and the two must never be copied into one another.",
"samples is the scorecard's auto_close_considered, not its corpus_items. A corpus of 72 items does not yield 72 closure decisions.",
"prompt_version and prompt_sha256 come from the prompt registry at the revision graded, not from the current tree.",
"corpus_digest and code_revision are copied verbatim from the scorecard so a row can be traced back to the run that produced it.",
"A row measured on a corpus that was tuned against is not independent and must not be promoted."
],
"records": []
}