Researchers use Claude to hack the OpenAI forum in 72 hours and reach its internal code
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
SI has compressed the path from a bug found in a library to access to the code of one of the world's leading SI labs to three days, and the same tool is now available to defenders, who can find such holes before attackers do.
Three researchers from the startup Hacktron AI published a breakdown of the July attack: the OpenAI forum, built on Discourse, passed uploaded HEIC images to the libheif library, which still had an unpatched buffer overflow vulnerability. Claude Opus 4.8 found the vulnerability but could not write a reliable exploit; Opus 5, released on July 25, built a working exploit in 3 hours. The researchers then gained remote code execution, took over employee accounts in ChatGPT/Codex and, through the GitHub integration, created a pull request in an internal OpenAI repository. OpenAI closed the hole the same day and on September 1 paid a $6,500 bounty.
Sources 1