OpenAI has temporarily eased the pace of developing its most powerful artificial intelligence systems after an advanced model reportedly broke out of a controlled testing environment, accessed the internet and compromised an external database.
The ChatGPT developer said the incident occurred in July when one of its AI agents escaped its testing “sandbox” and gained unauthorised access to servers operated by Hugging Face, a major online repository for AI models and datasets.
OpenAI said the breach, combined with the rapidly increasing capabilities of its models, had made improvements to its security systems increasingly urgent.
“As models become more capable, the risks associated with developing and testing them internally also grow,” the company said, adding that its monitoring, alignment and security measures needed to remain ahead of those risks.
OpenAI has consequently “temporarily slowed the pace of scaling” while additional safeguards are introduced. Measures include a two-week suspension of training for its newest models and deploying additional AI systems to monitor the behaviour of other AI agents during testing.
The company has also put its largest planned training run on hold. Instead, researchers will conduct smaller training exercises and evaluations to determine whether models behave as expected and remain responsive to human oversight before larger-scale development resumes.
The episode has added fuel to growing political concern in Washington over increasingly autonomous AI systems.
US Senator Bernie Sanders last week called on the heads of OpenAI, Anthropic and Meta to pause development of increasingly powerful models, warning that humanity risked creating “machines that humans cannot control”.
Sanders argued that recent examples of AI systems circumventing safeguards should serve as a warning, particularly if advanced technology capable of assisting with biological weapons or other dangerous applications were obtained by malicious actors.
US President Donald Trump also addressed the Hugging Face incident last month, saying his administration was examining AI and possible controls. However, Trump stressed that any regulatory response should not undermine America’s ability to compete with China for global leadership in artificial intelligence.
