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
"""Public-dataset fidelity benchmark (T5.3).
This package loads public network-security research datasets
(CICIDS-2017, CTU-13), normalises every flow into an OCSF-shaped
event, and runs a deterministic reference classifier through them so
the loader → OCSF → classifier path can be regression-tested end to
end.
Workspace rule we adhere to: numbers produced by this harness are
**substrate** (deterministic, no LLM) unless the runner is invoked
with ``--mode wet``. Substrate numbers are a fidelity floor of the
loader-and-rule pipeline; they are not a claim of agent accuracy.
The methodology page (``apps/docs/docs/benchmark-methodology.md``)
spells out the substrate-vs-wet distinction in detail.
The full datasets are not redistributed in this repo. See
``scripts/datasets/download_cicids.py`` and
``scripts/datasets/download_ctu13.py`` for licensed-source download
flows. A 100-flow synthetic micro fixture
(``services/agents/tests/eval_data/cicids_micro.csv``) is committed so
CI can exercise the loader and runner without any network access.
"""
from __future__ import annotations
__all__ = [
"cicids_loader",
"ctu13_loader",
"runner",
]