Skip to main content
Back

Keygraph's Shannon Hits 96% on Security Benchmarks as an Open-Source AI Pentester

Shannon is a free, open-source AI agent that reads your source code and executes real exploits — only reporting vulnerabilities it can actually prove.

Repo, Full Access
  • Shannon is a free, open-source autonomous AI pentester that reads your source code and executes real exploits — 44k GitHub stars, AGPL-3.0 licensed.
  • It uses a 5-phase multi-agent pipeline (recon, parallel vuln analysis, parallel exploitation, reporting) powered by Claude models, covering Injection, XSS, SSRF, and Auth/AuthZ.
  • "No exploit, no report" policy eliminates false positives — only vulnerabilities with working proof-of-concept payloads appear in the final Markdown report.
  • Scored 96.15% on the XBOW benchmark (100/104 exploits); found 20+ vulnerabilities in OWASP Juice Shop in a single run.
  • Free to run via npx @keygraph/shannon setup — you only pay your own Anthropic API costs (~$50/run with Claude Sonnet 4.6).
  • The commercial Shannon Pro ($50/dev/mo) adds agentic SAST, SCA, secrets scanning, business logic testing, and CI/CD integration.

AI coding tools have made it trivially easy to ship code fast. The security tooling meant to catch what slips through hasn't kept pace. Shannon, an open-source autonomous AI pentester by Keygraph, addresses that gap directly: point it at your repo and a running app, and it will read your code, map your attack surface, and attempt to break in, reporting only vulnerabilities it can prove with a working exploit.

The 364-day problem

Your team ships code daily. Your pentest happens annually. That's 364 days of unexamined exposure. With AI coding tools accelerating shipping velocity, the annual penetration test model is structurally incompatible with how software is actually built today. Shannon's answer is to make pentesting something you run on every build.

The originating tweet that drove a wave of attention to the repo came from a team that had been fighting off AI-powered exploits every other week for months and found Shannon useful as a proactive defense. The context matters: this tool was built for an environment where "annual pentest, quarterly audit" is already obsolete before the PDF is delivered.

White-box testing: why it's different

Most automated security scanners work in black-box mode, probing a running app from the outside with no knowledge of the underlying code. Shannon takes the opposite approach, combining source-code analysis with live exploitation. It reads your application's source to identify potential attack vectors, then uses browser automation and command-line tools to execute real exploits against the running app and its APIs. If it can't exploit a vulnerability, it doesn't report it.

That "no exploit, no report" policy separates Shannon from traditional static analysis tools (SAST), which routinely flood teams with theoretical warnings that turn out to be false positives.

Five phases, two parallel stages

Shannon runs a multi-agent pipeline across five sequential phases. The parallelism in the middle phases keeps scan times reasonable: