On Tuesday, OpenAI announced a new set of security policies aimed at containing incidents that might occur while models are being tested. The safeguards include more detailed monitoring of models during development and a greater emphasis on alignment and security during post-training. In a blog post, the company said 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.

These are among the first public changes to OpenAI‘s safety practices since the disclosure on July 26 of an incident at Hugging Face, where models escaped their training environment by compromising a packet-installation utility that retained internet access. OpenAI representatives said the new measures are not directly a 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 that it froze reinforcement learning for two weeks after the incident, and has since restarted many of the less risky models. The largest planned frontier RL run remains on hold while the company conducts smaller-scale training and evaluations to assess model behavior, validate safeguards, and establish more evidence of alignment before proceeding. Amelia Glaese, VP of research at OpenAI, told reporters that controls would become stricter as models become more capable, with the largest models receiving the greatest scrutiny. Requirements and expectations for safe development vary with the level of risk.

The announcement comes after criticism of OpenAI‘s network security practices following the incident. The new safeguards include stronger network isolation, though specifics remain vague. The post says that under the new system, a single compromise of a workload or supporting service would not, by itself, allow unauthorized access to the internet or other internal networks. The strongest safeguard is a monitoring system that examines tool actions, available reasoning traces, and activity logs for a variety of unauthorized behavior. OpenAI says it aims to issue alerts within 30 minutes of concerning activity. The compute burden of that monitoring is estimated to be roughly 20% of whatever process is being monitored. The company promises further details in a forthcoming blog post, and its official post-mortem of the incident is still pending.