luozhonglzw/cybersec-agent

Эта страница переведена автоматически. Технические детали смотрите в оригинале. Открыть оригинал

0 stars 0 forks Качество 58 Тренд 14MIT

Security operations AI agent built with LangGraph, deterministic policy enforcement, human approval, auditable tool use, and MCP.

Обзор

A security-operations agent prototype built on LangGraph, with a deliberate boundary between untrusted model output and authoritative planning, policy, and audit behavior. Security analysts spend much of their time turning raw telemetry into a defensible conclusion: is this indicator malicious, how serious is it, what should be done about it, and who approved that action. CyberSec Agent is a prototype that answers those questions over structured security events and threat-intelligence records — and that makes the answer auditable. The design premise is that the language model is an untrusted component. The model contributes intent, tool selection, and prose. The risk level, the response plan, the approval requirement, and the audit trail are produced by deterministic rule engines that never read the model's output. The evaluation harness is built to check that this separation actually holds. This is a research and engineering prototype.

README

Показан фрагмент оригинального README. Открыть оригинал.

CyberSec Agent

Release v0.5.0 · MIT License · Python 3.12+ · LangGraph · FastAPI

A security-operations agent prototype built on LangGraph, with a deliberate boundary between untrusted model output and authoritative planning, policy, and audit behavior.

中文项目介绍与面试材料 — 项目定位、技术栈、技术贡献、限制与面试问答（简体中文）。

Current release: v0.5.0 (2026-10-24) — knowledge-base Q&A over a version-controlled corpus, with three explicitly selected retrieval backends and structural citation validation.

Run the offline demo (no LLM provider required)

uv run python scripts/demo_offline.py

Four scenarios, each computed from a real return value — no API key, no network, no Docker. First time here? Quick Start covers environment setup; this section deliberately does not repeat it.

Scenario	Kind	What it demonstrates
`DEMO 1` RBAC	real HTTP through the real FastAPI app	401 / 403; unknown and missing credentials are byte-identical
`DEMO 2` Citation validation	real corpus, real retrieval, real validator	fabricated citations → `grounded=false`
`DEMO 3` Retrieval fail-closed	real Qdrant Local assembly path	broken index → `status="unavailable"`, no fallback
`DEMO 4` HITL	test-based (real pytest subprocess)	pause / resume / thread isolation / approver constraints

DEMO 2’s answer text is scripted — there is no provider in this run — but the retrieval, the evidence envelope and the citation validator are real code. DEMO 4 is a test-based demonstration, not a live /triage → /resume round trip: that path enters the graph at the agent node, so it needs a reachable provider. The script exits non-zero if any scenario fails.

Output of a real local run at b3e7950:

  PASS  DEMO 1  RBAC — authentication and role admission
  PASS  DEMO 2  Citation validation — fabricated citations are rejected
  PASS  DEMO 3  Retrieval fail-closed — a broken index is refused, not replaced
  PASS  DEMO 4  HITL approval and resume — TEST-BASED DEMONSTRATION

DEMO PASSED: 4/4 scenarios

Verified by CI

Job	Scope	Observed at `b3e7950`
`test`	offline suite — no network, no API key	2,283 passed, then the 4/4 demo
`test-postgres`	PostgreSQL integration against a throwaway server	373 passed, 0 skipped
`docker`	real image build + corpus load + route check	image built; 40 docs / 80 chunks in-container