OpenAI Employees' Accounts Compromised Through Image Exploit and Sign-In Flaw

Three security researchers leveraged Anthropic's Claude Opus 5 to build an image-processing exploit that, when chained with a critical weakness in OpenAI's authentication system, granted them access to employee ChatGPT accounts and internal code repositories in under 72 hours. The attack unfolded across two linked vulnerabilities: a remote code execution flaw in the libheif image library underlying OpenAI's community forum, and an overly permissive sign-in token system that granted forum credentials full API access to connected ChatGPT and Codex accounts.