Shannon runs a penetration test of a web application and its APIs on its own. It reads the application's source code to map architecture, trust boundaries, interfaces, data flows, and high-risk assets, derives candidate attack paths, and then drives browser automation and command-line tools against the running application to try to exploit them. A candidate that cannot be demonstrated with a working proof of concept is discarded rather than reported.
A run moves through six stages: reconnaissance and vulnerability analysis over the live application, agentic security code analysis, reconciliation that merges and deduplicates both streams into an exploitation queue, exploitation agents, validation, and reporting. Each scan runs in an ephemeral Docker container with the target repository mounted read-only and an isolated local workspace, so source code and model traffic stay inside your own infrastructure.
You supply the model credentials. Anthropic, OpenAI, xAI, and AWS Bedrock are configured directly by provider ID, any other provider in the Pi harness catalogue works with the same provider-and-model naming, and a custom base URL points any of them at a proxy or gateway. Local models served by Ollama, vLLM, or LM Studio work through their OpenAI-compatible endpoints, as do routers such as OpenRouter and gateways such as LiteLLM.
Reports come out as PDF and Markdown with evidence attached, plus JSON and SARIF 2.1.0, which is written by default on exploit-mode scans. Official GitHub Actions and GitLab CI/CD integrations run the same pipeline against staging or development targets, keep reports and logs as artifacts, and can fail a pipeline when a finding at or above a severity threshold is exploited; a code-analysis hypothesis alone never fails it.
Documented limits: exploitation agents mutate the target, so runs belong on sandboxed, staging, or local environments rather than production; a full run takes roughly 1 to 1.5 hours and incurs model costs; findings still need human review because model-generated reports can carry weakly supported details; and source code read by the agent can carry prompt injection, so untrusted codebases are out of scope.
Features
- Exploit-gated reporting: a vulnerability is reported only with a reproducible proof of concept, and unproven hypotheses are dropped
- Security code analysis: maps architecture, trust boundaries, interfaces, dependencies, data flows, and critical assets before exploitation begins
- Authenticated testing: credentials, login flows, TOTP, email authentication, focus areas, and rules of engagement come from configuration
- OWASP-focused coverage: specialised agents for injection, cross-site scripting, server-side request forgery, broken authentication, and broken authorization
- CI/CD gates: GitHub Action and GitLab component runs with artifact retention, code-scanning upload, and severity thresholds
- Report formats: PDF, Markdown, JSON, and SARIF 2.1.0 for any SARIF consumer
- Provider choice: built-in Anthropic, OpenAI, xAI, and AWS Bedrock, plus custom base URLs, local models, routers, and gateways under your own key
- Resumable workspaces: an interrupted scan resumes without repeating completed work
