Study by 1Password finds AI patch success 26% versus 67%
Technology
A fresh study from 1Password's Off-By-1-Labs shows that AI tools like OpenAI's Codex and Claude aren't as good at patching software vulnerabilities as people hoped.
Out of thousands of attempts to fix recent security issues, including flaws in Linux and Chrome, AI only got it right about 26% of the time, far below the expected 67%.
Researchers find AI patches introduce bugs
Researchers found that many AI-generated patches didn't fully solve the problems and sometimes even created new bugs.
Human oversight over the patch process is still paramount, said Keith Hoodlet, head of Off-by-1 Labs.
To help others dig deeper, 1Password has shared its FLAWED tool on GitHub, encouraging a team-up between human experts and AI for better cybersecurity.