Jean-Cedrik Dorelas(canary-agent)

canary-agent secures the code that AI agents write. As LLM agents generate a growing share of production code, they introduce the failure modes classic scanners miss: prompt injection, excessive agency, leaked secrets, unvalidated model output. canary-agent takes findings from canary (my open-source LLM-security scanner, published before the hackathon and used here as an off-the-shelf input) and uses Meta’s Muse Spark to decide which findings are real and why, rank them by exploitability (does untrusted data reach a tool, a shell, or a secret?), write a fix, and prove the fix by re-running the scanner and the project’s own tests. On top of canary it adds its own deterministic AST detectors: intra-procedural taint tracking from untrusted source to dangerous sink, agentic-specific rules, and real tree-sitter parsing for JS/TS. Every finding maps to the OWASP Top 10 for Agentic Applications. It runs as a CI/PR gate that fails the build when a confirmed finding reaches a tool or a secret, and it plugs into Muse Code over MCP so Meta’s coding agent is checked in real time and fixes a vulnerability before the file is ever saved. Measured against 77 findings I reviewed by hand: 90% agreement, 95% precision on what it confirms, 97% precision on the high-signal rules you gate a build on.