A team of three researchers from the security firm Hacktron successfully used Anthropic's Claude Opus 5 to exploit vulnerabilities and gain unauthorized access to OpenAI employee accounts and an internal code repository. This was achieved by chaining two distinct flaws: a bug in OpenAI's public help forum and a weakness in OpenAI's login system. This was part of a controlled security research effort, not a malicious attack. Hacktron reported these vulnerabilities to OpenAI, demonstrating their access with a harmless pull request. OpenAI responded promptly by fixing the issue within 14 hours and awarded Hacktron a $6,500 bounty. OpenAI confirmed the login flaw but did not provide detailed descriptions of the account takeovers. The research team emphasized that the issue lay within OpenAI's identity management system, which allowed access to multiple services through a single sign-on option shared with the public forum. The initial breach was caused by a vulnerability in the libheif library, which is used to handle HEIC and HEIF image files. This vulnerability allowed remote code execution on the Discourse-based forum server. Although this flaw was fixed in May 2026, the Discourse server used by OpenAI had not yet applied the patch. The researchers leveraged AI to expedite the process of crafting a working exploit, highlighting how AI models can significantly reduce the time and expertise needed for such tasks. Hacktron's wider research project, HEIF Heist, identified similar vulnerabilities in other major companies' software, although the extent of these findings remains partially unconfirmed. This case underscores the importance of promptly updating software and carefully managing single sign-on systems to prevent similar breaches.