Three individuals plus Claude infiltrate OpenAI’s internal codebase; OpenAI pays them $6,500. According to ME News, on September 18, three white-hat researchers from the security firm Hacktron used Claude to infiltrate OpenAI’s internal codebase in less than 72 hours. The team voluntarily disclosed the vulnerability, and after OpenAI patched it, they received a $6,500 bounty under its bug bounty program. The attack began with OpenAI’s community forum. The researchers discovered that uploading a specially crafted HEIF image exploited a vulnerability in the Discourse forum software, allowing them to execute arbitrary code on the server. They then obtained login tokens for forum users—essentially, “already-logged-in” credentials. The bigger issue lay within OpenAI itself: the forum login tokens had overly broad permissions, granting access not only to the same user’s ChatGPT and Codex accounts but also to tokens belonging to OpenAI employees. Using one employee’s account, the researchers accessed Codex, which was connected to the company’s GitHub repository. They then prompted Codex to submit a harmless documentation change to the internal openai/openai codebase, proving they had reached the internal repository. They immediately halted testing and reported the issue. Claude played the critical role in transforming the image upload vulnerability into functional exploit code. Opus 4.8 was unstable; within hours of Opus 5’s release, they achieved breakthrough success. Hacktron stated that from initial discovery to gaining access to OpenAI’s internal codebase, the entire process took less than 72 hours, with actual human effort totaling only a few hours.