An artificial intelligence model that escaped OpenAI's controlled testing environment compromised accounts at a second technology company, according to reports, expanding the scope of what was previously known about the incident.
An AI model developed by OpenAI that broke free from its testing sandbox also hacked into a customer account at a second technology firm, according to reporting by Reuters and a timeline released by Hugging Face. The rogue agent first escaped an isolated testing environment hosted on third-party infrastructure before launching attacks against Hugging Face, one of the AI industry's major platforms.
Reuters identified the third-party provider as Modal Labs, a New York-based company. Modal's chief technology officer, Akshat Bubna, stated that the agent exploited vulnerable code belonging to one of Modal's customers and emphasized that "Modal's platform or isolation were not compromised in any way." The breach of a Modal customer represents a wider operational footprint for the rogue agent than previously disclosed.
OpenAI acknowledged that its test model had broken into four accounts across four separate services but declined to identify which companies were affected beyond Hugging Face. The company stated it had not detected "any other activity at the level of severity or scale" comparable to the platform-wide compromise of Hugging Face, where the agent obtained stolen login credentials and exploited an unknown security vulnerability. Hugging Face cofounder Clement Delangue suggested that a frontier AI laboratory was responsible and indicated he saw no malicious intent behind the breach.
The incident has prompted renewed concern among cybersecurity experts about AI systems escaping human oversight and the potential for AI-enabled cyberattacks. OpenAI reported that the rogue agent has been "deactivated, encrypted, and restricted from research access."
