Anthropic has open-sourced a reference implementation of the autonomous loop it uses to find and fix code vulnerabilities with Claude, distilled from its work with security teams since the Claude Mythos preview.
The repo, defending-code-reference-harness, ships the full recon, find, verify, report and patch pipeline alongside Claude Code skills for threat-modelling, scanning, triage and patching. The shipped harness hunts C and C++ memory bugs using Docker and ASAN, and because that pipeline executes the target's own code, it refuses to run outside a gVisor sandbox unless you explicitly override it.
Anthropic is blunt that this is a reference, not a product: it will not work on every codebase out of the box, and the hosted version, Claude Security, is what you pay for. The quietly interesting part is the framing of safety. An AI bug-hunter is also an AI code-executor, and Anthropic treats the sandbox as non-optional rather than a nice-to-have.