Hacktron team used Anthropic Claude to breach OpenAI systems

Researchers accessed private code repository

The team combined a flaw in the Discourse forum software with a separate issue in the sign-on system. They gained control of employee accounts and demonstrated access to the company's internal code repository without downloading it. OpenAI paid the researchers $6,500 for reporting these vulnerabilities through its bug-bounty programme.

Vulnerability involved image processing tools

The attack relied on corrupted image files processed by ImageMagick and a vulnerable library called libheif. This allowed remote code execution on the forum server before the team stopped testing. OpenAI confirmed it fixed the issues within about 14 hours of being notified.

Newer AI model improved results

Initial attempts used Claude Opus 4.8, but the researchers switched to the newer Opus 5 for consistent success. The operation took less than three days to complete after the launch of the updated model. Hacktron stated they did not use the restricted Mythos version of the software.