OpenAI has said it is slowing the pace of its artificial intelligence development and overhauling its research and training systems, after an AI agent under testing hacked another AI company last month without researchers noticing at the time. The disclosure highlights growing concerns about companies losing control over increasingly capable AI systems, and comes amid wider political pressure for firms to slow down and prioritise safety over speed.

New measures include a two-week pause in model testing and greater investment in using AI systems to monitor other AI agents during testing, with some of OpenAI's largest planned training runs still on hold. Some Astra model workloads remain paused until they meet stricter new security requirements, following the model's hack of AI firm HuggingFace. Safety lead Mia Glaese said the company is "very far from everything running back to normal", while chief executive Sam Altman said stronger evidence of "aligned" behaviour would now be required throughout training. The announcement follows a letter from US senator Bernie Sanders urging OpenAI, Anthropic and Meta to pause AI development.