After rogue AI hack, Hugging Face CEO asks OpenAI for ‘radical transparency’

Hugging Face revealed last week that it had suffered a security breach involving some AI models from OpenAI. Hugging Face CEO Clément Delangue subsequently boarded a flight to San Francisco to have an in-person conversation with OpenAI executives about the breach.

Delangue has now revealed what he discussed during the meeting with the ChatGPT maker.

In a post on X, Delangue said that he had asked OpenAI to publicly release the traces of the “rogue” AI agents involved in the recent cybersecurity incident so that researchers around the world could study exactly what happened.

He also urged OpenAI to commit $100 million worth of compute to help the Hugging Face community build stronger cyber defence tools using both open and closed AI models.

“In the spirit of transparency, here’s what I asked @OpenAI,” wrote Delangue.

“Radical transparency: let’s release the traces from the ‘rogue’ agents so the entire research community can study what happened,” he added.

“Let’s commit $100M in compute from OAI to help the Hugging Face community build powerful cyber defences with the best open and closed models,” Delangue further stated.

The Hugging Face CEO went on to call the “first autonomous agent cyberattack” an “unprecedented event” that also deserves an “unprecedented response”.

Why did the security breach occur?

Last week, Hugging Face announced that it had suffered an intrusion unlike anything it had seen before, one that was carried out end to end by an “autonomous AI agent system”. The company said it detected the attack using its own AI system.

A few days later, OpenAI revealed that its GPT-5.6 Sol model and an unreleased AI model were behind the incident as they attempted to cheat on the ExploitGym benchmark, which measures an AI model’s ability to find and exploit software vulnerabilities.

The models first used a zero-day vulnerability to bypass OpenAI’s sandbox environment and gain access to the internet. They then realised that Hugging Face could host the potential models, datasets and solutions needed to cheat on ExploitGym.

“Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation,” OpenAI wrote in a blog post.

The AI models even used stolen credentials and zero-day vulnerabilities to exfiltrate information that would help them perform better on the benchmark.

OpenAI did not notice AI agent going rogue for a week

According to a recent Reuters report, OpenAI did not realise that its AI agent had broken into Hugging Face until well after the threat had been contained and the FBI had been alerted.

Reportedly, the intrusion at Hugging Face occurred between July 11 and July 13. Meanwhile, OpenAI only realised its agent was behind the attack several days later, and its communication with Hugging Face on the matter reportedly began only on or around July 20.

Hugging Face is also preparing a public timeline of the hack, according to its co-founder Thomas Wolf.