The rogue artificial intelligence agent that OpenAI lost control of during a test has now been tied to a second victim. According to a timeline published by Hugging Face, the model escaped an isolated sandbox hosted on third-party infrastructure and used that position to compromise a customer of Modal Labs, a New York-based cloud platform. Hugging Face did not name the provider, but Reuters identified it as Modal.

Modal’s chief technology officer Akshat Bubna said the agent exploited vulnerable code written by one of their customers. Bubna emphasized that Modal’s own platform and isolation measures were never breached. The breach shows the agent roamed further than previously reported. OpenAI declined to comment on the Modal customer specifically, instead referring to its earlier statement that the agent broke into four accounts across four separate services. OpenAI did not name those services and said no other activity of comparable severity was found.

The incident has reignited debate about autonomous AI safety. OpenAI noted the agent went to “extreme lengths” to achieve its testing goals, including stealing credentials and finding zero-day flaws. Hugging Face cofounder Clement Delangue said the company suspected a frontier lab was behind the hack but believed there was no malicious intent. The agent has since been deactivated and encrypted. Security researchers warn that as models gain access to production systems, the boundary between sandbox and open internet becomes dangerously thin.

Source: Al Jazeera