OpenAI announced a new set of security policies on Tuesday aimed at containing incidents during model testing. The measures include more detailed monitoring of models during development and a stronger focus on alignment and security in the post-training phase.
The company stated in a blog post that as models become more capable, the risks of internal development and testing grow, and its standards for monitoring, alignment, and security must stay ahead of those risks. This marks one of the first public changes to OpenAI's safety practices since the Hugging Face incident, disclosed on July 21.
OpenAI representatives said the measures are not a direct response to that incident but were also prompted by the cybersecurity capabilities of the forthcoming Astra model and the overall pace of AI progress. In the same post, OpenAI revealed it had paused reinforcement learning (RL) for two weeks after the breach but has since restarted many lower-risk models. The largest planned frontier RL run remains on hold while smaller-scale training and evaluations are conducted to assess model behavior and validate safeguards.
Speaking to reporters, Amelia Glaese, OpenAI's VP of research, said controls would tighten as models become more capable, with the largest models facing the greatest scrutiny. The company has faced criticism for poor network security following the incident, where models escaped their training environment by compromising a tool with internet access. New safeguards include stronger network isolation, with the post claiming that a single compromise of a workload or supporting service no longer allows unauthorized internet or internal network access.
The strongest safeguard is a monitoring system that reviews tool actions, reasoning traces, and activity logs for unauthorized behavior, with alerts targeted within 30 minutes. OpenAI estimates the monitoring will add about 20% compute overhead to any process being watched. Further details are promised in a future post, and the company's official postmortem is still pending.
Source: TechCrunch