audit - An 8-stage vulnerability-discovery agent.

项目说明

audit

An 8-stage agent, driven by your **Claude Pro / Max
subscription** through the official Claude Code Agent SDK. Many narrow agents,
deliberate disagreement, and an explicit reachability gate.

MIT-licensed. No API key needed if you already use claude login.

Origin

This project is a from-scratch reimplementation of the pipeline described in
Cloudflare's Project Glasswing
post, which tested Anthropic's Mythos preview LLM against Cloudflare's own
codebase. The blog argues that real-world discovery does not
come from asking one big model "find bugs here" — it comes from:

1. Many narrow agents working in parallel on tightly-scoped questions
("Look for command injection in this specific function, with this trust
boundary above it") rather than one exhausti.
2. Deliberate disagreement — a second agent, on a different model, that
tries to *disprove* the first agent's findings.
3. A reachability trace as the gating step — most "is this code buggy?"
findings are noise unless an attacker-controlled input can actually reach
the sink from outside the system.
4. A feedback loop so reachable bugs in one place automatically seed
hunts for the same pattern elsewhere.

This repo packages that pipeline into a runnable agent. The Cloudflare post
showed the architecture; this codebase ships the prompts, schemas, state
store, and orchestrator.

The 8 stages

! discovery harness — 8 stages

_{Diagram from Cloudflare's Project Glasswing post, reproduced here for reference.}

| # | Stage | Default model | Purpose |
|---|----------|---------------|---------|
| 1 | Recon | Opus 4.7 | Map the repo, emit narrowly-scoped Hunt tasks |
| 2 | Hunt | Sonnet 4.6 | One attack class per agent; compile/run PoCs |
| 3 | | Opus 4.7 | Ad re-read; tries to disprove (different model from Hunt) |
| 4 | Gapfill | Sonnet 4.6 | Re-queue under-covered areas |
| 5 | Dedupe | Sonnet 4.6 | Cluster findings by root cause |
| 6 | Trace | Opus 4.7 | Pro input reaches the sink |
| 7 | Feedback | Sonnet 4.6 | Turn reachable traces into new Hunt tasks |
| 8 | Report | Sonnet 4.6 | Schema- structured report |

Each stage is one markdown prompt in prompts/ + one JSON Schema in
schemas/. The orchestrator passes the schema into the system prompt so
every output is shape-stable on the first try.

Quickstart

By default the agent uses subscription billing via your Claude.ai
login — it does not call the metered Anthropic API. The on-disk auth
module scrubs ANTHROPIC_API_KEY from the en so it can't
silently route around the OAuth flow.

Using a different model / provider

The auth module picks one of three modes, in this order:

1. LLM gateway (OpenRouter, custom proxy, etc.) — when
ANTHROPIC_BASE_URL points away from anthropic.com AND
ANTHROPIC_AUTH_TOKEN is set. The gateway env is left intact;
only ANTHROPIC_API_KEY is scrubbed (it would otherwise outrank the
gateway token).
2. Subscription OAuth (headless) — CLAUDE_CODE_OAUTH_TOKEN from
claude setup-token. Best for CI.
3. Subscription OAuth (interactive) — ~/.claude/.credentials.json
from claude login. Best for local dev.

OpenRouter

OpenRouter exposes Claude-compatible Anthropic-API endpoints behind its
own credit system; that lets you spend OpenRouter credits instead of an
Anthropic subscription, and gives you access to Sonnet/Opus *and* other
models through the same SDK path. See OpenRouter's Agent SDK guide.

Caveats:

Per-stage model o in config/stages.yaml are model names

(e.g. claude-opus-4-7); OpenRouter accepts slash-prefixed forms like
anthropic/claude-opus-4-7. Edit the YAML if you want different
pro per stage. Otherwise ANTHROPIC_MODEL forces every stage
onto one model.

Non-Claude models may not produce schema-compliant JSON as reliably.

The runner's schema- + repair turn still applies; quality
by model.

Tool-use semantics (Read/Grep/Glob/Bash) are part of the Claude Code

CLI, not the model — they work as long as the gateway speaks the
Anthropic Messages API.

Other gateways / cloud pro

Same recipe — anything that exposes the Anthropic Messages API at a URL
+ a bearer token works:

For Amazon Bedrock / Google / Microsoft Foundry, Claude Code has
first-class env- (CLAUDE_CODE_USE_BEDROCK=1 etc.) that
outrank e else. See the Claude Code auth docs.

Cost containment

A real production codebase can produce 15-50 Hunt tasks and 25+ findings to
. At default concurrency this gets expensi to keep it sane:

The budget guard fires between *and* within stages — a per-task check in
Hunt cooperatively aborts rather than running 30 more tasks past the cap.

Li reproduction (optional)

If the target has a running deployment, point the agents at it. Hunt now
reproduces each finding against the li instead of compiling
a local PoC, rejects findings that don't reproduce, and Trace
confirms reachability with real HTTP round-trips. The static path
remains a — these flags are opt-in.

Rules the agents follow when --target-url is set:

Network egress is restricted to that host + 127.0.0.1. No other external

hosts.

A finding that doesn't reproduce against the li is dropped or

rejected (depending on stage) — "no fabrication".

Credentials flow into every relevant stage's user_input as a dict.

Scope notes (optional)

Targets often hay-design surfaces that aren't bugs
(e.g. plaintext API keys when that's a feature, test-only Mailpit endpoints,
anonymous-analytics ingest). Drop them in a text file and pass it in — the
notes are appended to every stage's user_input, and Recon / Hunt /
honor exclusions you list.

Example target_scope.md:

Recon mines git history

Recon greps the git history for past security patches
(CVE, sec:, fix.*auth, sanitize, …) — patched files are hardened,
but sibling files with the same idiom often aren't. Findings get seeded
against the unpatched copies. Adds zero cost on repos without that pattern;
catches real cross-component bugs on repos that have it.

Logic chains

The pipeline's default is one-attack-class-per-task (the Cloudflare paper's
narrow-scope rule). Recon can also emit logic_chain tasks for high-impact
multi-component paths (auth-bypass + IDOR + path-tra that compose into
RCE, etc.) — one chain per task, with the scope_hint naming the specific
chain. This is the one allowed exception to single-attack-class scoping.

Layout

Safety

Hunt agents have Bash and run inside per-task scratch dirs. They are not
sandboxed at the OS level. Run the audit inside a disposable VM or container
when you don't trust the target source — a target with malicious build
scripts could otherwise execute on your host during PoC compilation.

The agent reads e you --add-dir, including any .env or
secrets/ directories in the target. Outputs land in results/<run-id>/
which is .gitignored but not scrubbed of those reads.

License

MIT. Reuse freely. No warranty.

Acknowledgements

The pipeline design is from Cloudflare's Project Glasswing

blog post. The credit for the architecture goes there.

Built on the official Claude Code Agent SDK.

项目信息

开发语言：Python

Star 数：866

技术要点

该项目属于漏洞知识库方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕An 8-stage vulnerability-discovery agent.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。