Hacktron AI uses Claude to breach OpenAI's GitHub in under 72 hours

Security researchers at Hacktron AI used Anthropic's Claude AI to breach OpenAI's private GitHub environment in under 72 hours, chaining two vulnerabilities to gain access. The attack began with a heap buffer overflow in libheif, triggered via HEIC image uploads on OpenAI's Discourse forum, which allowed remote code execution. Claude Opus 4.8 initially failed to create a reliable exploit due to address-space layout randomization, but the newer Claude Opus 5 succeeded within hours. The three-person team switched to Claude Opus 5 after its July 24 release, which produced a working exploit quickly. The breach highlights AI's growing capability in offensive cybersecurity, though details on the second vulnerability and OpenAI's response remain undisclosed.