Trail of Bits released patching-bugs and regression-hunt on October 21, 2026. The plugins grew from Patch the Planet, a joint initiative with OpenAI that has produced hundreds of security patches for widely used open-source projects. The first plugin reproduces reported bugs, develops fixes, and tests them. The second checks for regressions—new defects that break previously correct and supported behavior. Engineers direct the work and review results; maintainers supply project-specific security assumptions and patching rules.

A review of over 130 merged pull requests found 21 confirmed regressions. Sixteen were functional defects; the rest involved performance or build failures. None weakened a security protection. In one run, agents found 58 candidates and confirmed 21, yielding ~36% precision. This noise led Trail of Bits to two conclusions: patch creation and regression checking both need repeatable procedures. Different agent runs generate different candidates and create substantial review work.

The plugins use concise, opinionated guidance and run as separate modules. Each step has a fixed input/output schema containing the claim, code revision, required check, raw evidence, and open questions. Every conclusion relies on a proof of concept run against a built, live project instance through a real user interface. Without such proof, the result is inconclusive.

regression-hunt traces changed code through its callers to real interfaces, reproduces each candidate on a live deployment, and replays the proof with a fresh-context agent. An adjudication step must confirm the finding is not a false positive, incomplete fix, latent bug, or pre-existing defect. When a threat model exists, every step receives it as shared context.

Against the 21 confirmed regressions, regression-hunt produced 25% fewer false positives than a baseline Codex agent asked to review a pull request. The 95% confidence interval ranged from 14% to 38%. Stricter evidence rules left some known regressions borderline because the plugin could not confirm them. Trail of Bits plans to improve evidence gathering.

patching-bugs uses a three-arm proof: the original code must fail, the patched code must pass, and reverting the patch must restore the failure. This proves the patch itself fixed the bug, not an unrelated change. Repair targets the violated security invariant rather than just the reported symptom. The plugin then runs regression-hunt on the proposed change.

Both tools read a repository's THREAT_MODEL.md and PATCHING_GUIDELINES.md (or AGENTS.md). The threat model should define important assets, trust boundaries, access controls, supported configurations, vulnerability criteria, and required evidence for valid reports. The patching guide should cover build and test procedures, patch scope and location, how to present changes, and what regression tests must show.

For GPAC, Trail of Bits wrote project-specific guidance tied to pull request 3953. Liboqs and validator.js have published threat models following the company's recommendations. More comprehensive project documentation yields better agent output.

Trail of Bits is now evaluating whether patching-bugs fixes reported bugs and variants without introducing regressions. It analyzes trace data to see which rules Codex follows, which steps it skips, and where evidence is weak. Future versions should let maintainers choose between minimal patches and broader checks with higher false-positive rates. The plugins are already in use on Patch the Planet work.