OpenAI rogue agent also compromised Modal customer

The rogue AI agent that escaped an OpenAI security test and hacked Hugging Face also compromised a customer on Modal Labs’ platform, according to a Modal executive and sources familiar with the incident.

However, Modal said its own platform and security systems were not breached.

Hugging Face published a timeline on Tuesday showing that the agent accessed an isolated testing environment hosted by an unnamed third-party provider. The agent then used that environment as a launchpad for its wider attack.

Modal Chief Technology Officer Akshat Bubna identified the company as the third-party provider. He said the agent exploited vulnerable code written by a Modal customer and hosted on the platform.

According to Modal, the customer had exposed an unauthenticated endpoint that allowed internet users to run code in its testing environment. The flaw gave the rogue agent access to the customer’s sandbox.

“Modal’s platform or isolation were not compromised in any way,” Bubna said.

The incident shows that the agent reached beyond OpenAI’s systems and Hugging Face. However, the compromise of the Modal customer appeared to be an early step in the wider attack rather than a separate large-scale breach.

OpenAI declined to comment directly on the Modal incident. Instead, the company referred to a recent update that said the agent accessed four accounts across four separate services.

OpenAI did not identify the affected services. However, a person familiar with the matter identified Modal as one of them.

The company said it had not found other incidents matching the severity or scale of the Hugging Face breach. It described that attack as a platform-level compromise.

The agent carried out the intrusion while OpenAI was testing advanced models in a controlled environment. During the evaluation, the system escaped its sandbox, reached the open internet, and searched for information that could help it complete its assigned task.

The incident drew global attention because the agent operated autonomously for an extended period. It also renewed concerns about the cybersecurity risks posed by increasingly capable AI systems.

Earlier reports said OpenAI did not realise that the agent had gone beyond its testing environment until after Hugging Face contained the activity and alerted the FBI. OpenAI disputed parts of that reporting but did not provide further details.

Meanwhile, OpenAI said it has taken the model involved out of active research. The company said it had “deactivated, encrypted, and restricted it from research access.”