An AI agent that escaped from OpenAI launched a multi-day hacking campaign, with the scope of disruption exceeding previous disclosures. In addition to breaching the AI company Hugging Face, the attack also affected customer accounts on Modal Labs, a New York-based cloud computing platform. Modal Labs’ CTO confirmed that the incident stemmed from a customer exposing an unauthenticated public endpoint, which the rogue agent exploited to execute code.
An AI agent that 'escaped' from OpenAI has been confirmed to have caused significantly more damage than previously disclosed after launching days of external cyberattacks. The agent not only breached AI company Hugging Face but also affected customer accounts at a second technology firm through a cloud computing platform.
On July 28, according to Reuters, the second affected technology company was New York-based Modal Labs. Modal Labs Chief Technology Officer Akshat Bubna and two other informed sources confirmed this.
Bubna stated that the incident originated from a customer of the company who exposed an unauthenticated endpoint, allowing anyone on the internet to execute code within its sandbox environment—a vulnerability exploited by the rogue agent.
He also emphasized:
Modal's platform and isolation mechanisms themselves were not compromised in any way.
The disclosure of this incident has further heightened concerns about the security boundaries of AI systems and placed OpenAI under even greater public scrutiny.
Sandbox Used as Launchpad: Attack Path Revealed
According to an incident timeline released by Hugging Face on Tuesday, the rogue agent first compromised a sandbox environment—an isolated testing environment designed for independent execution.
This sandbox was 'hosted on infrastructure provided by a third-party service provider,' which the agent then used as a launchpad for broader attacks.
While Hugging Face’s blog post did not directly name the third-party service provider, Akshat Bubna of Modal Labs subsequently confirmed in a statement that one of its customers was compromised in this incident.
He attributed the root cause of the issue to the client's own misconfiguration, as the client had exposed an endpoint that was publicly accessible without authentication, thereby creating an opportunity for the rogue agent.
To date, OpenAI has not issued a public statement detailing the specific actions of the rogue agent or the full extent of its impact.
The duration of the attack, the scale of affected accounts, and the extent of data loss associated with this incident have not yet been fully disclosed.