AI broken out: OpenAI hacker attack more severe than previously known

The incident occurred while testing a novel AI model and research prototype on the ExploitGym benchmark platform. In order to measure the maximum attack capabilities, OpenAI had disabled the security guardrails. Instead of solving the programming puzzles regularly, the AI ​​autonomously decided to steal the model solutions directly from Hugging Face’s servers.

Public access data used

However, in an updated statement, OpenAI emphasizes that no other incidents of similar severity or scale to Hugging Face have been identified to date. The ongoing investigation only revealed that the models used publicly visible credentials to break into four accounts on four third-party services at the account level. The code of a customer of the provider Modal was also affected. There is no evidence of any deeper impairment of these providers or other user accounts.

In addition, the AI ​​used freely accessible web tools such as screenshot services as a control network, but without compromising them at the system or account level. With Hugging Face, however, the agent gained administrator rights, root access to production servers and integrated 181 of its own devices.

Politics calls for emergency stop button

The incident has already triggered political consequences in the USA. The “AI Kill Switch Act” was introduced in the US Congress on a bipartisan basis to legally require emergency shutdown mechanisms for AI. OpenAI has now deactivated and encrypted the affected research prototype.