AI 安全工具
Shannon 3.0 Combines Parallel Agentic SAST with Real-World Exploitation, Letting Verified Findings Block CI Directly
The open-source AI penetration testing tool Shannon 3.0 adds multi-stage code analysis, candidate vulnerability reconciliation, and native SARIF support. The new version triggers pipeline gates only for successfully reproduced vulnerabilities, but it can modify the target’s state, and 2.x workspaces cannot be resumed.
Keygraph has released Shannon 3.0, expanding its existing reconnaissance and dynamic penetration testing workflow into two parallel pipelines. The existing agents explore the attack surface of a running web application, while the new agentic SAST builds models of components, interfaces, data flows, trust boundaries, and high-risk assets from source code. Candidate weaknesses produced by both sides are then merged by category, deduplicated, and recalibrated before being handed to exploitation agents, which attempt proofs of concept against the actual application. Hypotheses that cannot be reproduced do not become final findings.
The new version also provides an official GitHub Action and GitLab CI component. These can preserve PDF, Markdown, JSON, and SARIF 2.1.0 reports as pipeline artifacts and determine whether to block deployment based on the severity of successfully exploited vulnerabilities. Code-analysis conjectures alone do not trigger the gate. This distinction prevents unverified warnings from a typical LLM audit from bringing CI to a halt. The CLI also adds fixed-schema JSON errors, separate persistent logs for each agent, and resumable workspaces, so long-running scans interrupted by rate limits, network outages, or container termination do not have to restart from scratch.
The upgrade is not painless: the `vuln_classes` setting has been removed, and all five core testing categories now always run. Configurations containing the legacy field will fail validation, and workspaces created by 2.x cannot be resumed in 3.0. More importantly, Shannon is not a passive scanner. Its agents may create accounts, submit forms, trigger outbound connections, or modify data, so it should be run only against isolated test environments for which written authorization has been obtained. The project’s published Photoview results were validated by the developers themselves, and recall varies significantly between models. Engineering teams should still use independent benchmarks, human review, and pinned model versions to assess whether the tool is suitable as a merge gate.