LLM-assisted vulnerability autofixing: approaches and how to validate the patches
At DARPA's AIxCC finals in August 2025, AI systems patched 68% of vulnerabilities they found — up from 25% at semifinals. Here's how the approaches differ and why validation still matters most.