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
fix(agents): fail closed when LLM triage errors on a severe alert (#1249)
antigravitysoham-eng156d24ca056c
diff --git a/services/agents/app/workers/fused_alert_consumer.py b/services/agents/app/workers/fused_alert_consumer.pyindex 21b66d999..b9a0af14b 100644--- a/services/agents/app/workers/fused_alert_consumer.py+++ b/services/agents/app/workers/fused_alert_consumer.py@@ -102,6 +102,7 @@ _METRICS = {"outcome_written": 0,"outcome_suppressed": 0,"ungrounded_demoted": 0,+ "llm_failed_held_for_review": 0,"persist_retries": 0,"dead_lettered": 0,"approvals_raised": 0,@@ -287,6 +288,36 @@ def build_state(message: dict[str, Any]) -> InvestigationState | None:)+# Severities a failed model call may not close. The deterministic scorer is a+# fallback for when the model is unreachable, and on a laptop CPU or a gateway+# with the wrong base URL that is every call. Observed: a critical+# "ransomware canary files encrypted" alert, vssadmin deleting shadow copies,+# came back `likely_benign` from the fallback after the LLM timed out.+_FAIL_CLOSED_SEVERITIES = frozenset({"critical", "high"})+++def _hold_severe_for_review_after_llm_failure(state: InvestigationState) -> InvestigationState:+ """Fail closed: an LLM error is never the reason a severe alert closes.++ Only auto-closeable verdicts on high or critical alerts change, and only on+ the path where the model was asked and failed. A deployment that runs+ deterministic triage on purpose keeps today's behaviour.+ """+ severity = str((state.raw_alert or {}).get("severity") or "").strip().lower()+ if severity not in _FAIL_CLOSED_SEVERITIES:+ return state+ verdict = state.verdict+ if normalize_disposition(str(verdict or ""), default=NEEDS_REVIEW) not in AUTO_CLOSEABLE_DISPOSITIONS:+ return state+ state.add_finding(+ f"LLM triage failed, so the deterministic verdict '{verdict}' on a {severity}-severity "+ "alert is held for review instead of closing it."+ )+ state.verdict = NEEDS_REVIEW+ _METRICS["llm_failed_held_for_review"] += 1+ return state++class FusedAlertTriageWorker:"""Consumes ``aisoc.alerts.fused`` and auto-triages each alert (copilot)."""@@ -1143,7 +1174,7 @@ class FusedAlertTriageWorker:logger.warning("auto_triage_worker.llm_failed_fallback", error=str(exc))state = await run_triage(state)_METRICS["deterministic"] += 1- return state, "deterministic"+ return _hold_severe_for_review_after_llm_failure(state), "deterministic"def _apply_groundedness_gate(self,diff --git a/services/agents/tests/test_alert_triggered_playbooks.py b/services/agents/tests/test_alert_triggered_playbooks.pyindex 4a96dd464..7278631a2 100644--- a/services/agents/tests/test_alert_triggered_playbooks.py+++ b/services/agents/tests/test_alert_triggered_playbooks.py@@ -209,14 +209,28 @@ class TestTheWiring:assert "alert_trigger.run_for_alert(" in source, "no playbook can start from an alert, which is the defect 5.1 exists to fix"def test_it_runs_after_triage_not_before(self) -> None:- """A playbook's conditions read the verdict and the confidence."""+ """A playbook's conditions read the verdict and the confidence.++ Located inside `triage` by AST rather than by `str.index` over the+ whole module. The substring form pinned whichever occurrence came+ first in the file, so adding *any* helper above `triage` that also+ reads `state.verdict` moved the anchor into the helper and failed a+ correctly-ordered worker — which is what a fail-closed LLM-error+ guard did. The property under test is the order of these statements+ within the function that runs them, so that is what this reads.+ """+ import astimport pathlib- source = pathlib.Path(__file__).resolve().parents[1].joinpath("app/workers/fused_alert_consumer.py").read_text()- trigger_at = source.index("alert_trigger.run_for_alert(")- verdict_at = source.index("verdict = state.verdict")- assert trigger_at < verdict_at- assert source.index("tokens = tracker.total_tokens") < trigger_at+ path = pathlib.Path(__file__).resolve().parents[1].joinpath("app/workers/fused_alert_consumer.py")+ tree = ast.parse(path.read_text())+ triage = next(n for n in ast.walk(tree) if isinstance(n, (ast.FunctionDef, ast.AsyncFunctionDef)) and n.name == "triage")+ body = "\n".join(path.read_text().splitlines()[triage.lineno - 1 : triage.end_lineno])++ trigger_at = body.index("alert_trigger.run_for_alert(")+ verdict_at = body.index("verdict = state.verdict")+ assert trigger_at < verdict_at, "the playbook trigger must run before the verdict is read back"+ assert body.index("tokens = tracker.total_tokens") < trigger_atclass TestTheTriggerEventMatchesTheCorpus:diff --git a/services/agents/tests/test_fused_alert_worker.py b/services/agents/tests/test_fused_alert_worker.pyindex f8b1e7638..61f74db5e 100644--- a/services/agents/tests/test_fused_alert_worker.py+++ b/services/agents/tests/test_fused_alert_worker.py@@ -321,3 +321,41 @@ class _FakeGovernor:def check(self, tenant_id, fingerprint): # noqa: ANN001, ARG002return self._decision+++async def _llm_raises_and_fallback_says(monkeypatch, verdict: str):+ monkeypatch.setattr(worker_mod, "resolve_llm_config", _fake_llm_config(allowed=True))+ monkeypatch.setenv("AISOC_DETERMINISTIC", "0")++ async def _raise(_state):+ raise RuntimeError("litellm.Timeout: Connection timed out. Timeout passed=60.0")++ async def _fallback(state):+ state.verdict = verdict+ return state++ monkeypatch.setattr(worker_mod, "run_auto_triage", _raise)+ monkeypatch.setattr(worker_mod, "run_triage", _fallback)++[email protected]("severity", ["critical", "high", "CRITICAL"])+async def test_llm_failure_never_closes_a_severe_alert(monkeypatch, severity):+ # A failed model call is not evidence. The fallback said likely_benign on a+ # critical ransomware alert once; it must be held for review instead.+ await _llm_raises_and_fallback_says(monkeypatch, "likely_benign")+ result = await _worker().triage(_fused(severity=severity))+ assert result["tier"] == "deterministic"+ assert result["verdict"] == "needs_review"++[email protected]("severity", ["low", "medium", "info"])+async def test_llm_failure_on_a_minor_alert_keeps_the_fallback_verdict(monkeypatch, severity):+ await _llm_raises_and_fallback_says(monkeypatch, "benign")+ result = await _worker().triage(_fused(severity=severity))+ assert result["verdict"] == "benign"+++async def test_llm_failure_does_not_downgrade_a_true_positive(monkeypatch):+ await _llm_raises_and_fallback_says(monkeypatch, "true_positive")+ result = await _worker().triage(_fused(severity="critical"))+ assert result["verdict"] == "true_positive"