1Password research finds AI models fail to properly patch vulnerabilities 74% of the time

According to 1Password's Off-By-1-Labs security research team, frontier AI models tasked with generating software patches achieve success rates significantly lower than the hypothesized 67%, with LLMs producing incomplete or incorrect fixes for complex vulnerabilities at high rates. The research contradicts assumptions that models trained on open source code and public disclosures would reliably generate working patches.