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
"""CrowdStrike detections must carry the acting identity through to fusion.
Falcon puts the user on each behavior, not on the detection, and the envelope
omitted it entirely. That is not a cosmetic gap: fusion's correlation key is
``{tenant}:{entity}:{tactic}``, so an EDR detection against a named user
arrived with the entity segment blank and correlated into the same bucket as
every other CrowdStrike alert in the tenant.
"""
from __future__ import annotations
from app.connectors.crowdstrike import CrowdStrikeConnector
def _connector() -> CrowdStrikeConnector:
return CrowdStrikeConnector(client_id="id", client_secret="secret")
def _detection(**overrides) -> dict:
detection = {
"detection_id": "ldt:abc:123",
"max_severity_displayname": "High",
"created_timestamp": "2026-09-23T10:00:00Z",
"device": {"hostname": "WKSTN-01", "external_ip": "203.0.113.7"},
"behaviors": [
{
"display_name": "Suspicious PowerShell",
"technique_id": "T1059.001",
"user_name": "alice",
"user_id": "S-1-5-21-1",
}
],
}
detection.update(overrides)
return detection
class TestActor:
def test_actor_comes_from_the_behavior(self):
event = _connector().normalize(_detection())
assert event["actor"] == "alice"
def test_first_behavior_with_a_user_wins(self):
event = _connector().normalize(
_detection(
behaviors=[
{"display_name": "Stage one", "user_name": ""},
{"display_name": "Stage two", "user_name": "bob"},
]
)
)
assert event["actor"] == "bob"
def test_missing_actor_is_none_not_an_empty_string(self):
# An empty string is a value the normalizer would map onto
# actor.user.name, producing an alert that claims to name a user and
# names nobody. None is absent, which is the truth.
event = _connector().normalize(_detection(behaviors=[{"display_name": "No user"}]))
assert event["actor"] is None
def test_no_behaviors_at_all_is_handled(self):
event = _connector().normalize(_detection(behaviors=[]))
assert event["actor"] is None
assert event["title"] == "CrowdStrike Detection"
class TestEnvelope:
def test_still_emits_the_canonical_envelope(self):
# `source` + `raw_event` is what makes the normalizer take the
# canonical path, which is what gives the event OCSF class 2001 and
# therefore makes it promotable to an alert at all.
event = _connector().normalize(_detection())
assert event["source"] == "crowdstrike"
assert isinstance(event["raw_event"], dict)
def test_severity_and_host_survive(self):
event = _connector().normalize(_detection())
assert event["severity"] == "high"
assert event["hostname"] == "WKSTN-01"
assert event["src_ip"] == "203.0.113.7"
assert event["mitre_techniques"] == ["T1059.001"]