tech π€ AI patching software is failing most of the time
A new study revealed that AI-generated software patches are largely ineffective, showing only a 26% success rate. The research by 1Password's Off-By-1-Labs tested LLMs against six complex, recently disclosed vulnerabilities. Across 6,080 patch attempts, LLMs failed to create a fix or introduced new bugs 53.9% of the time. The core issue is that AI produces 'Fix-Like Artifacts with Embedded Defects' instead of solid fixes. Human oversight remains crucial for determining which bugs need patching and when. π€