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
-- Record how well each auto-triage verdict was supported by its evidence.
--
-- `app.confidence.groundedness` scores what fraction of the concrete
-- indicators a verdict's reasoning asserts — IPs, hashes, CVEs, MITRE
-- techniques, domains — actually appear in the evidence the agent was given.
-- v8.0 wired it into the triage path so an auto-closing verdict below the
-- floor is demoted to `needs_review` instead of closing.
--
-- The score itself was not persisted. It survived only inside a findings
-- string, which means it could not be aggregated, trended, or used to answer
-- the question a buyer evaluating an AI-SOC actually asks: how often is your
-- agent's reasoning supported by what it was shown, and how often does it
-- decline to decide?
--
-- Both columns are nullable rather than defaulted. NULL means "not scored",
-- which is a different fact from "scored zero" — pre-migration rows and
-- verdicts from the deterministic path were never assessed, and defaulting
-- them to 0.0 would read as every historical verdict being unsupported.
ALTER TABLE alerts
-- Fraction in [0, 1]. NULL = not scored.
ADD COLUMN IF NOT EXISTS triage_groundedness DOUBLE PRECISION,
-- True when the groundedness gate demoted an auto-closing verdict to
-- needs_review. Distinct from `disposition = 'needs_review'`, which is also
-- reached by low confidence and by genuine escalation: this flags the subset
-- where the agent had a confident answer that its evidence did not support.
ADD COLUMN IF NOT EXISTS triage_ungrounded BOOLEAN;
-- The abstention and groundedness aggregates on /metrics/funnel scan by
-- tenant and window, so index the shape those queries use. Partial, because
-- the vast majority of rows are never scored and including them would make
-- the index mostly dead weight.
CREATE INDEX IF NOT EXISTS ix_alerts_tenant_groundedness
ON alerts (tenant_id, created_at)
WHERE triage_groundedness IS NOT NULL;
COMMENT ON COLUMN alerts.triage_groundedness IS
'Fraction of indicators cited in the triage reasoning that appear in the evidence. NULL = not scored.';
COMMENT ON COLUMN alerts.triage_ungrounded IS
'True when the groundedness gate demoted a confident auto-closing verdict to needs_review.';