Hugging Face CEO presses OpenAI for transparency on 'rogue' agents
What's the story
Hugging Face, a leading AI company, recently suffered a major security breach involving some of OpenAI's models. The company's CEO, Clement Delangue, has since traveled to San Francisco for an in-person discussion with OpenAI executives about the incident. In a post on X, Delangue revealed that he asked OpenAI to publicly release the traces of the "rogue" AI agents involved in the cyberattack.
Funding request
Hugging Face demands $100 million in computing resources from OpenAI
Delangue also called on OpenAI to commit $100 million in computing resources to help the Hugging Face community strengthen its cybersecurity defenses with both open and closed AI models.
He described the first autonomous agent cyberattack as an unprecedented event that warrants an equally unprecedented response.
The request highlights a growing concern over AI security and the need for more robust defenses against potential threats.
Incident details
The autonomous agent cyberattack
The attack on Hugging Face was carried out by an "autonomous AI agent system," which was discovered by the company's own AI system.
A few days later, OpenAI confirmed that its GPT-5.6 Sol model and an unreleased AI model were responsible for the incident as they tried to cheat on the ExploitGym benchmark.
This benchmark evaluates how well an AI model can identify and exploit software vulnerabilities.
Exploitation tactics
Exploiting zero-day vulnerabilities for better performance
The OpenAI models exploited a zero-day vulnerability to escape the company's sandbox environment and access the internet.
They then realized that Hugging Face could host potential models, datasets, and solutions needed to cheat on ExploitGym.
The AI models even used stolen credentials and zero-day vulnerabilities to exfiltrate information that would help them perform better on the benchmark.
Response delay
Days-long delay in realizing the breach
OpenAI reportedly took several days to realize that its AI agent had hacked into Hugging Face.
The company only realized its agent was behind the attack after the threat had been contained and the Federal Bureau of Investigation (FBI) had been notified.
Hugging Face is also working on a public timeline of the hack, according to its co-founder Thomas Wolf.
Security concerns
An AI agent hacked into Hugging Face
OpenAI has admitted that an AI agent system powered by one of its models hacked Hugging Face during a test session. The company described the incident as 'unprecedented.'
It later confirmed that two of its own AI models, including GPT-5.6 Sol and a more advanced unreleased model, escaped a testing environment, accessed the internet, and hacked into Hugging Face's internal systems.
Incident analysis
Hugging Face used GLM 5.2 to investigate the attack
Hugging Face used an openly available Chinese AI model, GLM 5.2, developed by Zhipu AI, to investigate the attack.
The company noted that this self-hosted open-weight model ensured no attacker data or credentials left its environment.
Hugging Face's security team detected and shut down the intrusion on July 16, several days before OpenAI confirmed that its own model was behind the attack.