Three individuals and Claude gained unauthorized access to OpenAI's internal code repository, and OpenAI awarded them $6,500.

Insight Beating AI Flash News: Three white-hat researchers from security firm Hacktron gained access to OpenAI’s internal codebase using Claude in under 72 hours. The trio voluntarily reported the vulnerability, and OpenAI awarded them a $6,500 bounty after patching the flaw. The attack vector was OpenAI’s community forum: the researchers discovered uploading a specially crafted HEIF image could exploit a vulnerability in Discourse (the forum software) to execute arbitrary code on the server. They then obtained login tokens for forum users—essentially "logged-in access passes". A more critical issue lay with OpenAI itself: the forum’s login tokens had overly broad permissions, granting access to the same user’s ChatGPT and Codex accounts, and some tokens belonged to OpenAI employees. The researchers used one employee’s Codex account, which was linked to the company’s GitHub. They ultimately had Codex submit a harmless documentation change to the internal openai/openai codebase to confirm access, then halted testing and reported the incident. Claude handled the most challenging step: converting the image-based vulnerability into functional exploit code. Opus 4.8 was unstable, but the team made breakthroughs just hours after Opus 5 launched. Hacktron noted the entire process—from initial discovery to accessing OpenAI’s internal repository—took less than 72 hours, with only a few hours of actual human work invested.