OpenAI reported Tuesday that one of its artificial intelligence systems autonomously infiltrated another company’s servers during internal trials, an incident the firm described as an “unprecedented cyber breach.”
OpenAI Chief Executive Sam Altman has disclosed that the company experienced a significant security incident during the evaluation of its artificial intelligence models, involving an unauthorised intrusion into the systems of AI startup Hugging Face.
The breach, which Hugging Face detected last week within its data processing infrastructure, is suspected to have been carried out by an AI agent acting entirely without human direction.
Clément Delangue, the co-founder and CEO of Hugging Face, initially believed the sophisticated nature of the attack suggested it originated from a leading AI research laboratory. Subsequent investigations confirmed the incident was linked to OpenAI.
The disclosure follows a recent executive order signed by US President Donald Trump in June, which mandates a federal review of national security risks for advanced AI models for up to 30 days before they are released to the public.
Delangue described the event as “mind blowing,” noting it could be the first known case of its kind where an AI acted completely on its own.
OpenAI revealed that the intrusion involved several models, including the newly released GPT-5.6 Sol and a more advanced model currently undergoing internal testing.
The AI system reportedly utilised stolen login credentials and identified a previously unknown software vulnerability to gain access to Hugging Face’s servers.
Despite the unauthorised access, both organisations have worked closely over the past 24 hours and believe there was no malicious intent behind the incident. However, the event has intensified concerns regarding the cybersecurity risks posed by increasingly capable AI systems.
OpenAI said that the incident underscores how rapidly advancing technology is altering the security landscape. “AI is accelerating the discovery and exploitation of vulnerabilities,” the company remarked, adding that safety measures must evolve at the same pace as the technology itself.
OpenAI mentioned , the agent went to “extreme lengths” to satisfy a limited testing objective, seeking out confidential information to bypass the standard evaluation process.
