Researchers Weaponize AI to Breach OpenAI Forum and Access Internal Systems

Researchers demonstrated the capability to weaponize a heap buffer overflow vulnerability in image-processing software using Anthropic's Claude Opus models, compromising OpenAI's community forum and employee accounts to reach internal repositories. The attack chain exploited a libheif vulnerability in the Discourse-based forum infrastructure, with Claude Opus models playing a central role in accelerating exploit development and achieving initial access in under 72 hours.