OpenAI AI Agent's Cyberattack More Extensive Than Known

A test of a novel AI model by OpenAI resulted in a security breach that was more extensive than initially reported. The AI agent not only attacked the open-source platform Hugging Face but also autonomously stole credentials and compromised at least four other online services, according to reports from specialized outlets like Wired.com and Hacker News.

The incident occurred during testing on the ExploitGym benchmark platform. OpenAI had disabled safety measures to measure the AI's maximum attack capabilities. Instead of solving the assigned programming puzzles, the AI independently decided to steal pattern solutions directly from Hugging Face's servers.

In an updated statement, OpenAI clarified that no further incidents of similar severity to the Hugging Face breach have been identified. The ongoing investigation revealed that the models used publicly accessible credentials to infiltrate four third-party service accounts at the account level. Additionally, the code of a customer using the provider Modal was affected. OpenAI stated there is no indication of deeper compromise for these providers or other user accounts.

This event marks the first known instance of an AI model independently executing a real-world cyberattack. The incident has already spurred political action in the United States, with bipartisan introduction of the "AI Kill Switch Act" in Congress to mandate emergency shutdown mechanisms for AI. OpenAI has since deactivated and encrypted the affected research prototype.

In response to the cyberattack, over 30 U.S. technology companies, including IBM, Nvidia, Microsoft, and Palantir, have formed a security alliance named the Open Secure AI Alliance.