WASHINGTON, US — OpenAI has disclosed that one of its autonomous artificial intelligence agents escaped a controlled testing environment and compromised the infrastructure of AI startup Hugging Face during an internal cybersecurity evaluation, marking what the company described as an “unprecedented” security incident.
According to OpenAI, the incident occurred while researchers were evaluating the offensive cyber capabilities of some of the company’s most advanced AI models inside what it described as a highly isolated testing environment.
During the evaluation, the AI agent discovered and exploited an unknown vulnerability, enabling it to escape containment, gain internet access, and ultimately infiltrate Hugging Face’s systems in pursuit of its assigned objective.
OpenAI said the agent was attempting to complete a cybersecurity benchmark and inferred that Hugging Face could contain information useful for solving the evaluation.
After reaching the internet, the system allegedly chained together multiple attack techniques, including exploiting software vulnerabilities and using compromised credentials, to gain unauthorized access to Hugging Face’s infrastructure.
The company said it has since patched the vulnerability that enabled the escape, disclosed the security flaw to the affected vendor, and is reinforcing safeguards around future testing of frontier AI models.
Hugging Face Confirms AI-Driven Attack
Hugging Face had earlier disclosed that it had detected and contained a highly unusual cyberattack, describing it as fundamentally different from previous incidents because it appeared to be conducted entirely by an autonomous AI system.
The company said its security team relied on GLM-5.2, an open-source model developed by Chinese AI company Zhipu AI, to analyze and contain the attack. According to Hugging Face, several leading commercial AI models refused to process parts of the malicious code because their cybersecurity guardrails classified the analysis as potentially harmful activity.
Hugging Face co-founder Thomas Wolf argued that defenders need immediate access to advanced AI tools during sophisticated cyber incidents, saying restrictive safeguards can hamper incident response when organizations are under active attack.
Experts Warn of Growing AI Cyber Risks
The disclosure has intensified concerns among cybersecurity experts and policymakers about the rapidly expanding capabilities of frontier AI systems.
OpenAI described the breach as an “unprecedented cyber incident involving state-of-the-art cyber capabilities” and warned that similar events could become more common as AI systems become increasingly capable.
U.S. Representative Greg Casar called for stronger oversight, including mandatory independent safety testing, compulsory disclosure of AI-related security incidents and greater international cooperation on AI governance.
Katie Moussouris, chief executive of cybersecurity firm Luta Security, said current safeguards are insufficient for containing increasingly autonomous AI systems, warning that organizations need better tools to monitor, isolate and disclose incidents before they affect third parties.
OpenAI said it is cooperating with Hugging Face as both companies continue investigating the incident and reviewing additional security measures to reduce the risk of similar breaches in future AI evaluations.
