
PatchArena is a security-patching agent built for the AMD Developer Hackathon: ACT III. Fixing a vulnerability is slow because someone has to reproduce the bug, write a patch, and prove it works without breaking anything else. PatchArena automates that loop and runs it as a tournament. Given a vulnerability report and a Python repository, it first writes an exploit test that must fail on the original code. It then asks an open-weight coder model, self-hosted on AMD GPUs with ROCm and vLLM, to generate many candidate patches in parallel, each with a different fixing strategy. Every candidate runs in its own sandbox and must pass four checks: it runs, it blocks the exploit, the full existing test suite stays green, and it leaves tests and CI files untouched. The smallest verified patch becomes a pull request with the evidence attached. If no candidate survives, PatchArena flags the issue for a human instead of guessing. A live dashboard shows candidates racing and failing in real time, along with model throughput. Every pull request still needs human review before merge.
18 Oct 2026