What is Shannon?
Shannon is an autonomous penetration-testing agent for web applications and their APIs, built by Keygraph. You give it a URL and the path to the source code. It maps the architecture, trust boundaries and data flows from the code, explores the running app, then sends exploitation agents to attempt proof-of-concept attacks through a browser and command-line tools. A finding goes in the report only if an exploit worked.
It covers five classes of flaw: injection, cross-site scripting, server-side request forgery, broken authentication and broken authorisation. It handles logins, TOTP and email authentication through configuration, resumes interrupted scans, and writes PDF, Markdown, JSON and SARIF reports. Official GitHub Action and GitLab CI components can fail a pipeline when an exploit succeeds above a chosen severity.
You bring the model. Anthropic, OpenAI, xAI and AWS Bedrock are built in, and other providers, gateways and local models served by Ollama, vLLM or LM Studio work through a custom configuration. Each scan runs in a throwaway Docker container that mounts your repository read-only. The open-source edition is AGPL-3.0; Keygraph sells hosted Pro and self-hosted Enterprise editions on an enhanced build of the same engine.
Is Shannon worth it?
Shannon is the first open-source tool that makes routine, exploit-backed pentesting of your own web app cheap enough to run per release. It reads the code to decide where to attack, attacks the running app, and drops anything it cannot prove, so the report is short and each item comes with a working proof of concept. About 51,400 GitHub stars in just over a year, and v3.5.0 on 27 October 2026, show how many teams want that.
The numbers need care, because they are Keygraph's own. The README used to lead with 96.15% (100 of 104 challenges) on what Keygraph called a hint-free, source-aware variant of XBOW's benchmark; a modified benchmark is not comparable with published XBOW scores, and the claim has since left the README. The current evidence is a comparison Keygraph ran itself against a Doyensec study sponsored by Aikido: on Photoview 2.4.0, Shannon with Claude Opus 5 reported 24 issues, 23 of them real, for $115 in tokens, while Aikido found 32 and XBOW 7 at $4,000 per scan. A DeepSeek v4 Flash run cost $6.10 and found 18. Doyensec did not verify Shannon's results.
The catch: it is an attacker, and it behaves like one. Its agents create users, submit forms, change or delete data and send outbound requests, so the docs say to run it only with explicit written authorisation from the system's owner, only against local, sandboxed or staging copies with disposable data, and never in production; a disposable virtual machine is the safest place. Pointing it at code you do not trust is a prompt-injection risk. Anthropic and OpenAI apply cyber safeguards that can stop a scan halfway unless your account is verified, Claude Code subscriptions stopped working after v1.9.0, a full run takes 1 to 1.5 hours by the docs (2h 24m to 5h 26m in the Photoview test), and no token count per run is published, only these dollar figures and an earlier estimate of about $50 with Claude Sonnet 4.5.
It is a different tool from PyRIT, garak and DeepTeam, which attack the language model or chatbot; Shannon attacks the web app around it and does not test prompts. Claude Security is closer: it also reads whole repositories for logic flaws, but it does not exploit them, and it needs Claude Enterprise. Shannon needs source code, covers five flaw classes, and leaves dependency, secrets and black-box testing to the paid Pro edition at $50 per active developer per month.
Our take: think of Shannon as a junior pentester who never gets tired and has to show its working. It is good value on staging, and a person still has to read each finding before it becomes a ticket.
Where it earns its place
A 10-developer SaaS team shipping weekly to a staging environment
Find exploitable injection and access-control bugs before each release
Add KeygraphHQ/shannon-action to the release workflow against staging, store the model key as a secret and fail on high severity. Model spend ran $6.10 to $115 per scan in Keygraph's test, set by the model you choose.
An AppSec lead at a company with 200 internal apps and one annual pentest
Cover the long tail of apps that never get a human test
Run Shannon with a cheap model on every app's staging copy and a stronger model on the riskiest ones, then send confirmed findings to human pentesters for triage. Get the owners' written authorisation for every target first.
A pre-Series-A fintech with 12 developers
Decide between the open-source build and Keygraph Pro
The free Community Program gives Pro, with black-box testing, SCA, secrets scanning and fix pull requests, to start-ups with up to 20 developers; after that Pro is $50 per active developer per month, still with your own model key.
Before you commit
- Setup
- Docker and Node.js 22 or later, then npx @keygraph/shannon@latest setup for keys and npx @keygraph/shannon@latest start -u <staging URL> -r <repo path>. Clear your provider's cyber-safeguard verification first or the run can fail partway.
- Integrations
- Anthropic, OpenAI (including ChatGPT Plus and Pro subscriptions), xAI and AWS Bedrock built in; other providers, LiteLLM or OpenRouter gateways and Ollama, vLLM or LM Studio through a custom base URL or models file. GitHub Action (private repositories), GitLab CI component, SARIF 2.1.0 output.
- Data and privacy
- Runs on your machine in an ephemeral container with the repository mounted read-only; Keygraph says it never receives your code or proxies model traffic, but source and app context go to the model provider you configure unless you use a local model. A Temporal web UI listens on localhost:8233. Results stay in a local workspace.
- Cost at scale
- The software is free; each scan bills model tokens, $6.10 to $115 on Photoview by Keygraph's figures, and no per-run token counts are published. Keygraph Pro is $50 per active developer per month, model costs still yours; Enterprise is custom.
- Limits
- White-box only: it needs the source. Five flaw classes; no dependency, configuration or business-logic findings in the open-source build; reports can contain wrong details and need human review; Keygraph accepts no outside code contributions; AGPL-3.0 applies if you offer it as a service.
What’s good
- Reports only findings with a working exploit
- Model costs from $6.10 to $115 per scan in Keygraph's test
- Any provider or local model, with your own key
- GitHub Action and GitLab CI gates with SARIF output
What’s not
- Mutates data; unsafe outside an authorised sandbox
- Benchmarks are Keygraph's own
- Only five flaw classes and needs source code
- Cyber safeguards can stop scans on Claude or OpenAI
Best for
Engineering teams with their own web app, its source and a staging copy who want exploit-backed security tests on every release without paying for a pentest each time.
Shannon alternatives
Shannon: common questions
- Is Shannon worth it?
- Shannon turns a code review into proven exploits for the price of a model bill, which makes it worth running on every staging release, and never anywhere you lack written permission to attack. We rate it 4.1 out of 5.
- How much does Shannon cost?
- Shannon is free under AGPL-3.0; you pay your own model costs. Keygraph Pro is $50 per active developer per month, free for qualifying US nonprofits and pre-Series-A start-ups with up to 20 developers; Enterprise is custom-priced.
- What are the downsides of Shannon?
- Mutates data; unsafe outside an authorised sandbox; benchmarks are Keygraph's own; only five flaw classes and needs source code; cyber safeguards can stop scans on Claude or OpenAI.
- Who is Shannon best for?
- Engineering teams with their own web app, its source and a staging copy who want exploit-backed security tests on every release without paying for a pentest each time.
- How much does Shannon cost for a team or in production?
- The software is free; each scan bills model tokens, $6.10 to $115 on Photoview by Keygraph's figures, and no per-run token counts are published. Keygraph Pro is $50 per active developer per month, model costs still yours; Enterprise is custom.
- Is Shannon safe for company data?
- Runs on your machine in an ephemeral container with the repository mounted read-only; Keygraph says it never receives your code or proxies model traffic, but source and app context go to the model provider you configure unless you use a local model. A Temporal web UI listens on localhost:8233. Results stay in a local workspace.
- What are the best Shannon alternatives?
- The closest alternatives we have reviewed are PyRIT, garak, DeepTeam and Trail of Bits Skills.
Answers checked Nov 7, 2026.
Make Shannon? Show this rating on your site
