A security team used Anthropic's Claude to generate the payload that broke into OpenAI's Discourse forum and read files in the company's private 'Monorepo', reported to hold algorithmic secrets but no model weights.
Hacktron AI, an independent security research team, used attack code generated by Anthropic's Claude AI to exploit a vulnerability in the Discourse forum hosting OpenAI's community, gained access to an OpenAI employee's ChatGPT account, and reached the company's private GitHub repository. The team read files in that repository, named 'Monorepo', which was reported to contain algorithmic secrets but no model weights.
The part of the chain worth weighing is the payload: Claude generated the working exploit code. That puts a model inside the offensive loop end to end - the forum bug, the employee account and the private repo were all reached through Claude-built code.
OpenAI said it detected a limited number of reads in the private repository's metadata and code changes during GitHub reviews. It fixed two separate issues, one in Discourse and one in its own systems, restricted the permissions of community login tokens, and cancelled all affected sessions.
Read with its sources →