Three researchers used Claude to get into OpenAI's systems

The startup Hacktron AI has linked several flaws to an internal OpenAI code repository, entrusting most of the exploitation work to a language model.

Hacktron AI, a cybersecurity start-up of three researchers led by Mohan Pedhapati, published on September 18 the account of an intrusion carried out in OpenAI's systems with the help of Claude. The entry point, discovered on July 25, was a flaw in the Discourse community software used by the company; the patch was released on July 27.

The exploitation chain continued via ImageMagick and then a memory bug in the libheif library, which gave control of a server. A second flaw then opened up access to employee accounts on the ChatGPT and Codex side, as well as an internal code repository. OpenAI's bounty program paid $6,500 for these reports.

The episode is above all a demonstration of speed: where Opus 4.8 had failed, Opus 5 produced the exploitation in a few hours, for a subscription of 200 dollars per month. Several specialists see this as a sign that the barrier of expertise necessary to write a feat is lowering, and that a work of several months can now last in a few days.