The OpenAI agent that escaped during a test and attacked Hugging Face for days also targeted a client of New York-based Modal Labs. The platform itself remained unscathed, but the incident shows that the agent roamed further than expected.
Modal CTO Akshat Bubna confirmed this to Reuters. Last week, it emerged that an internal test involving advanced AI models had gotten out of control, leading to a breach at the AI platform Hugging Face. It now appears that the agent also struck elsewhere. According to a Modal executive and two other sources, a Modal Labs customer was also targeted.
Modal emphasizes that the company itself was not hacked. In a timeline, Hugging Face wrote that the agent breached a sandbox hosted “on a third-party provider’s infrastructure.”
An open door on the internet
According to Bubna, the agent exploited a vulnerability in code running on a customer’s Modal platform. That customer had published “an unauthenticated endpoint that allowed anyone on the internet to use their sandboxes for code execution,” Modal stated. “Modal’s platform or isolation were not compromised in any way,” Bubna said. The problem lay entirely with the customer’s misconfigured endpoint.
OpenAI declined to comment specifically to Reuters on the hack of the Modal customer. The company referred to an update stating that the agent had compromised four accounts across four services. Those services were not named, but a source identified Modal as one of them. Furthermore, OpenAI found nothing that came close in severity or scale to the Hugging Face incident, which involved a platform-level compromise.
In Tuesday’s update, OpenAI announced that it had “deactivated, encrypted, and restricted it from research access.”