Researchers from Hacktron AI, including Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini, successfully breached OpenAI systems using code generated by Anthropic’s Claude model. The entire operation was completed in under 72 hours.
The attack began by exploiting an image upload vulnerability in the Discourse software used by OpenAI’s community forum. A memory-corruption flaw allowed the team to bypass safety filters, with Claude generating the necessary exploit code after earlier attempts failed.
Access to OpenAI’s private source code was achieved through a secondary flaw in the company’s single sign-on system. By hijacking an employee session, the researchers gained entry to the internal repository.
OpenAI awarded the team a $6,500 bounty for the discovery. The company patched the authentication vulnerability within 14 hours of the report, while Discourse addressed the image processing bug with a severity rating of 8.8 out of 10.
