KeygraphHQ/shannon

What it is

Shannon (open-source edition, version 3.0) — Keygraph's autonomous AI pentester that reads a target's source, derives attack paths, and executes real exploits against the running app/API, reporting only vulnerabilities with a reproducible proof of concept. The same agent also powers Keygraph's commercial platform; this repo is the standalone CLI you run yourself.

How it works

npx pulls the worker image from Docker Hub, starts local infrastructure, mounts the target repository read-only inside an ephemeral worker container, and writes results to a local workspace. It maps architecture, trust boundaries, interfaces, data flows, and critical assets first, then hands credible attack paths to the live exploitation agents (browser automation + CLI tools).

Preconditions / Affected

• Docker for the worker container; Node.js 18+ for the npx workflow.
• AI provider credentials (Anthropic, OpenAI, xAI, AWS Bedrock, or any harness-catalogue provider). Bring your own key; Keygraph never proxies model traffic. Custom base URL / LLM-gateway and custom model configs are supported.
• Provider cyber safeguards must be cleared first. Anthropic and OpenAI apply real-time safeguards to cyber-security workloads that can interrupt a scan mid-run; complete their "legitimate security tester" guidance before your first run.

What to do

npx @keygraph/shannon@latest setup

npx @keygraph/shannon@latest start \
  -u https://your-app.com \
  -r /path/to/your/repo

Subscription (vs API-credit) support differs by provider and is the key version gotcha:

• OpenAI Codex — ChatGPT Plus/Pro subscriptions supported on the latest Shannon.
• xAI (Grok) — subscriptions supported on the latest Shannon.
• Claude Code — subscriptions not supported on the latest Shannon. To use a Claude Code subscription you must pin version 1.9.0, the final release built on the Claude Agent SDK.

Caveats

• Only run against systems you own or have explicit written authorization to test. Do not run against production — Shannon actively executes exploits.
• Positioned as a complement to human pentesters, not a replacement; intended to cover the long tail of apps between annual expert assessments, run against staging/releases.
• Published results are from Open Source scans of Photoview 2.4.0 (the app in Doyensec's Aikido-vs-XBOW comparison); Keygraph evaluated its own results separately — check their follow-up for methodology, costs, and limitations before trusting the numbers.