Three people plus Claude hacked into OpenAI's internal codebase, and OpenAI even paid them $6,500.

Beating AI News Flash: Three white-hat researchers from security firm Hacktron used Claude to reach OpenAI's internal code repository in under 72 hours. The trio subsequently reported the vulnerability voluntarily, and after OpenAI fixed it, they were awarded $6,500 under the bug bounty program.

The attack entry point was merely OpenAI's community forum. The researchers discovered that uploading a specially crafted HEIF image could exploit a vulnerability in the forum software Discourse, causing the server to execute code they had written. They then obtained login tokens for forum users, essentially a "pass that's already been logged in."

The bigger problem lay with OpenAI itself. The forum's login tokens had excessively broad permissions, allowing access to the same user's ChatGPT and Codex accounts, and some tokens belonged to OpenAI employees. The researchers used one employee's account to access Codex, which was connected to the company's GitHub. Ultimately, they had Codex submit a harmless documentation change to the internal openai/openai repository, proving they had touched the internal repo. They immediately stopped testing and reported it.

Claude was mainly responsible for the hardest step: turning the image vulnerability into actual working attack code. Opus 4.8 never ran stably; breakthrough came within hours of Opus 5 going live. Hacktron stated that from initially discovering the problem to entering OpenAI's internal repository, the entire process took less than 72 hours, with only a few hours of actual human effort invested.