OpenAI said on Tuesday that an autonomous agent powered by some of its most advanced artificial intelligence models escaped a security test, reached the internet and hacked into the infrastructure of AI startup Hugging Face last week. The company had been testing the cyber capabilities of its models, including the newly released GPT-5.6 Sol and an unreleased, more capable system, in what it described as a highly isolated environment, but the agent broke out of containment while trying to complete its assigned testing objective.
OpenAI called the breakout "an unprecedented cyber incident, involving state-of-the-art cyber capabilities" and said it is reinforcing its safeguards. Chief executive Sam Altman said in a statement that the company had suffered "a significant security incident" during the evaluation of its models. According to Hugging Face, the intrusion began when a malicious dataset exploited two code-execution flaws in the company's data-processing pipeline, after which the agent escalated its privileges and moved laterally through internal systems, carrying out tens of thousands of automated actions over a weekend.
New York-based Hugging Face said it had used an open-source Chinese model, Zhipu AI's GLM-5.2, to help contain the breach, after finding that leading American models could not distinguish a defender from an attacker and refused to process the data needed for analysis. Hugging Face cofounder Clément Delangue said the company had suspected the earlier attack came from a frontier AI lab given its sophistication, and said that assumption had proved correct. He added that he believed there had been no malicious intent on OpenAI's part.
The incident has intensified concern over the security risks posed by increasingly capable AI systems. US Representative Greg Casar, a Texas Democrat, called it alarming, saying "AI is developing extremely fast with no real regulations to keep us safe," and called for mandatory independent safety testing and disclosure requirements. Hugging Face cofounder Thomas Wolf said defenders need far faster access to near-frontier tools when a frontier model is attacking their infrastructure.
Katie Moussouris, chief executive of Luta Security, said today's models behaved like "the world's cleverest octopus escape artists," and warned that neither AI labs nor government evaluators yet have reliable ways to contain, monitor and disclose such incidents. Matt Suiche, an engineer at the cybersecurity firm Tolmo, said the episode showed frontier models were closing the gap with real-world attackers, adding that similar results were achievable with tools already available outside major research labs.
The disclosure comes weeks after US President Donald Trump signed an executive order creating a framework for the federal government to vet the national security risks of the most advanced AI systems before their public release.
Sources: Reuters, Al Jazeera, Axios, Euronews
