Shannon：只报告能证明的漏洞——Keygraph 的 AI 渗透测试 Agent

by Mycelium Protocol

---

使用前提：Shannon 是一个真实执行漏洞利用的工具。只能在你拥有或有书面授权的系统上运行。Anthropic 和 OpenAI 对网络安全工作负载应用实时防护，可能中断扫描；首次运行前需完成各平台针对合法安全测试人员的认证流程。

---

渗透测试有一个根本性的速度问题：你的团队每天都在发布代码，但渗透测试一年只做一次。这意味着在其余 364 天里，你可能每次发布都在把漏洞推上生产环境。

Shannon 是 Keygraph 对这个问题的答案：一个可以在每次构建或发布时按需运行的自主 AI 渗透测试 Agent。

GitHub: https://github.com/KeygraphHQ/shannon | ⭐ 46,621 | AGPL-3.0 | TypeScript | 2025-09-27

---

核心主张：证明式漏洞报告

Shannon 的设计出发点是拒绝输出「推测性告警」。

传统 SAST 工具会告诉你「这里可能有 SQL 注入」——因为代码模式看起来可疑。Shannon 做的是另一件事：真正发起攻击，只有攻击成功了，这个漏洞才会出现在报告里，并附上完整的可复现 PoC 步骤。

这意味着报告里的每一条发现都是被证明可利用的，而不是需要人工再次验证的候选项。

---

多 Agent 架构

Pre-Reconnaissance（源码扫描）
         ↓
Reconnaissance（攻击面映射）
         ↓
     ┌───┴───┬───────┐
     ↓       ↓       ↓
 Vuln     Vuln    ...
(Injection) (XSS)
     ↓       ↓       ↓
 Exploit  Exploit  ...
(Injection) (XSS)
     └───┬───┴───────┘
         ↓
     Reporting

五个阶段：

Pre-Reconnaissance：从仓库源码识别框架、入口点、数据流和攻击面，在接触运行应用之前就完成初步分析。

Reconnaissance：探索运行中的应用，把运行时行为和源码上下文关联起来，绘制具体的攻击面地图。

并行漏洞分析：针对不同漏洞类型的专用 Agent 同时运行——Injection、XSS、SSRF、Broken Authentication、Broken Authorization。

Exploitation：真正发起 PoC 攻击。无法证明的假设直接丢弃，不进入报告。

Reporting：把验证过的发现、证据和修复建议汇总成 Markdown 报告。

每次扫描在一个独立的 Docker 容器里运行，有隔离的工作区，支持中断后恢复（--resume）。

---

快速上手

# 交互式配置向导（API key、模型、目标等）
npx @keygraph/shannon setup

# 对有源码的目标发起渗透测试
npx @keygraph/shannon start -u https://your-app.com -r /path/to/your-repo

Shannon 自动从 Docker Hub 拉取 worker 镜像，启动本地基础设施，把目标仓库以只读方式挂载进 worker 容器，输出结果到本地工作区。

支持的 AI 提供商：Claude（官方推荐）、OpenAI（含 ChatGPT Plus/Pro 订阅）、xAI、AWS Bedrock，以及任何 OpenAI 协议兼容端点。

前置要求：

• Docker（worker 容器必需）
• Node.js 18+
• 至少一个 AI 提供商的凭证
• Anthropic/OpenAI 合规认证：两个平台对网络安全工作负载应用实时防护，可能中断扫描。首次运行前完成各自针对合法安全测试人员的认证流程。

---

实测样本报告

目标	发现
OWASP Juice Shop	20+ 漏洞，包含认证绕过、SQL 注入、IDOR、SSRF
c{api}tal API	~15 个严重/高危 API 漏洞，包含命令注入、认证绕过、Mass Assignment
OWASP crAPI	15+ 严重/高危发现，覆盖 JWT、注入、SSRF、API 鉴权路径

以上报告均来自故意设计为有漏洞的靶场应用，可在仓库 sample-reports/ 目录查看完整内容。

---

关键能力

认证测试：配置文件可描述登录流程、测试凭据、TOTP、基于邮件的登录，以及渗透测试范围和交战规则。

可恢复工作区：扫描中断后无需从头重来，可以从断点继续。

白盒攻击规划：用源码分析引导动态测试，聚焦现实的攻击路径，而不是盲目模糊测试。

---

重要限制

Shannon Open Source 有几个需要明确的边界：

• 聚焦可利用漏洞：Injection、XSS、SSRF、Broken Authentication、Broken Authorization。更广泛的静态分析（依赖漏洞、配置问题）在商业版 Keygraph Platform 里。
• 发现仍需人工审核：LLM 生成的报告可能包含支持不足或不准确的细节。
• 不要扫描不可信代码库：读取源码的 AI 工具可能被提示注入攻击。
• 单次扫描约 1-1.5 小时，会产生 LLM API 调用费用。

---

Shannon 2.0

仓库最新的置顶公告是 Shannon 2.0 正式发布，2026 年 8 月正式宣布。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Shannon: An AI Pentester That Only Reports What It Can Prove

by Mycelium Protocol

---

Safety requirement: Shannon actively executes exploits. Only run it against systems you own or have explicit written authorization to test. Anthropic and OpenAI apply real-time safeguards to cybersecurity workloads that can interrupt a scan. Complete their guidance for legitimate security testers before your first run.

---

Penetration testing has a fundamental speed mismatch: your team ships code daily, but your pentest happens once a year. That leaves 364 days where you could be pushing vulnerabilities to production with every release.

Shannon is Keygraph’s answer: an autonomous AI pentesting agent that can run on demand against every build or release.

GitHub: https://github.com/KeygraphHQ/shannon | ⭐ 46,621 | AGPL-3.0 | TypeScript | 2025-09-27

---

Core Claim: Proof-by-Exploitation Reports

Shannon is built around rejecting speculative alerts.

A traditional SAST tool tells you “there might be SQL injection here” — because the code pattern looks suspicious. Shannon does something different: it actually executes the attack. Only if the attack succeeds does the finding appear in the report, with a complete reproducible PoC.

Every finding in the output has been proven exploitable. None are candidates that need manual verification.

---

Multi-Agent Architecture

Pre-Reconnaissance (source code scan)
         ↓
Reconnaissance (attack surface mapping)
         ↓
     ┌───┴───┬───────┐
     ↓       ↓       ↓
 Vuln     Vuln    ...
(Injection) (XSS)
     ↓       ↓       ↓
 Exploit  Exploit  ...
     └───┬───┴───────┘
         ↓
     Reporting

Five stages:

Pre-Recon: identifies frameworks, entry points, data flows, and attack surfaces from the source repository before touching the live application.

Reconnaissance: explores the running application, correlating runtime behavior with code-level context.

Parallel vulnerability analysis: specialized agents for Injection, XSS, SSRF, Broken Auth, and Broken Authorization run concurrently.

Exploitation: real proof-of-concept attacks. Hypotheses that can’t be proven are discarded, not reported.

Reporting: validated findings with evidence and remediation guidance, compiled into a Markdown report.

Each scan runs in an ephemeral Docker container with an isolated workspace. Interrupted scans are resumable.

---

Quick Start

# Interactive setup wizard
npx @keygraph/shannon setup

# Run a pentest against a source-available target
npx @keygraph/shannon start -u https://your-app.com -r /path/to/your-repo

Shannon pulls the worker image from Docker Hub, starts required local infrastructure, mounts the target repo read-only inside an ephemeral worker container, and writes results to a local workspace.

Supported providers: Claude (officially recommended), OpenAI (including ChatGPT Plus/Pro subscriptions), xAI, AWS Bedrock, and any OpenAI-protocol-compatible endpoint.

Requirements:

• Docker (for the worker container)
• Node.js 18+
• Credentials for at least one AI provider
• Provider compliance: Anthropic and OpenAI apply real-time safeguards to cybersecurity workloads. Complete their guidance for legitimate security testers before the first run.

---

Sample Reports

Target	Findings
OWASP Juice Shop	20+ vulnerabilities including auth bypass, SQL injection, IDOR, SSRF
c{api}tal API	~15 critical/high API findings including command injection, auth bypass, mass assignment
OWASP crAPI	15+ critical/high across JWT, injection, SSRF, and API authorization paths

All reports were produced against intentionally vulnerable applications. Full reports are in sample-reports/ in the repository.

---

Key Capabilities

Authenticated testing: configuration files describe login flows, test credentials, TOTP, email-based flows, focus areas, and rules of engagement.

Resumable workspaces: interrupted scans resume from where they stopped — no re-running completed agents.

White-box attack planning: source analysis guides dynamic testing toward realistic attack paths instead of blind fuzzing.

---

Limitations

• Covers actively exploitable issues: Injection, XSS, SSRF, Broken Auth, Broken Authorization. Broader static analysis (vulnerable dependencies, insecure config) is in the Keygraph Platform.
• Findings require human review. LLM-generated reports can contain weakly supported details.
• Do not scan untrusted or adversarial codebases — tools that read source are exposed to prompt injection.
• A full scan takes roughly 1–1.5 hours and incurs LLM API costs.

---

Shannon 2.0

The repository’s current pinned announcement is Shannon 2.0 officially released.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.