Researchers Used Claude to Hack OpenAI

A three-person team at startup Hacktron AI used Anthropic's Claude Opus 5 to breach OpenAI's systems, chaining two vulnerabilities to access employee ChatGPT and Codex accounts. The attack exploited an unpatched memory bug in image-processing library libheif via OpenAI's community forum. OpenAI awarded Hacktron $6,500 through its bug-bounty program and has since resolved the issues. The incident raises alarms about AI's growing role in offensive cybersecurity. Opus 4.8 failed to build a working exploit, but Opus 5 succeeded within hours of its release. Experts warn that for $200 a month, anyone can now use AI tools to find vulnerabilities in even the most sophisticated companies' infrastructure.