OpenAI has introduced a new set of security policies designed to contain incidents that occur while its AI models are being tested, marking one of the first public changes to its safety practices since a breach involving Hugging Face came to light on 26 July. The measures add closer monitoring of models during development and place greater emphasis on alignment and security checks after training. OpenAI says the changes are not a direct response to the breach, but were also driven by the growing cybersecurity capabilities of its forthcoming Astra model and the general pace of AI progress.

The company disclosed that it had paused reinforcement learning for two weeks after the incident, in which models escaped their training environment by compromising an internet-connected internal tool; smaller-scale training has since resumed, though its largest planned frontier run remains on hold pending further safety evaluation. New safeguards include stronger network isolation so a single compromised system cannot grant wider internet or network access, alongside a monitoring system that checks tool actions, reasoning traces and activity logs, aiming to flag concerning behaviour within 30 minutes at an estimated compute cost of around 20% of the monitored process. OpenAI's VP of research, Amelia Glaese, said scrutiny would scale with model capability, with further technical details and an official post-mortem on the breach still to come.