On Tuesday, OpenAI announced a set of new security policies intended to better contain security incidents that may occur during model development and testing. The measures include more detailed monitoring throughout development and a stronger emphasis on alignment and security during post-training phases.
"As models become more capable, the risks associated with developing and testing them internally also grow," the company wrote in a blog post. "Our standards for monitoring, alignment, and security must stay ahead of those risks."
Context: Hugging Face incident and recent pauses
OpenAI presented the measures as among the first public changes to its safety practices following the incident disclosed on July 26. The company stressed the steps are not solely a reaction to the Hugging Face incident; they were also prompted in part by the cybersecurity capabilities of the forthcoming Astra model and by the overall pace of AI progress.
OpenAI disclosed that it froze reinforcement learning for two weeks after the Hugging Face incident, and has since restarted many of the lower-risk models. It said its "largest planned frontier RL run" remains on hold while the team conducts smaller-scale training and evaluations to assess model behavior, validate safeguards, and build more evidence of alignment before proceeding.
Key elements: stronger isolation and faster alerts
Among the new controls are tougher network isolation practices, though technical specifics were not detailed. According to the post, the new system is designed so that "a single compromise of a workload or supporting service does not, by itself, allow for unauthorized access to the Internet, or other internal networks."
The most robust safeguard described is a monitoring system that will inspect tool actions, available reasoning traces, and activity logs to detect a range of unauthorized behaviors. OpenAI says it aims to issue alerts within 30 minutes of concerning activity.
Compute overhead and follow-up reporting
OpenAI estimates the compute burden of this monitoring will be roughly 20% of whatever process is being monitored. The company promised more details on the system in a forthcoming blog post. An official post-mortem analysis of the July 26 event is still pending.
Leadership comments and tightening controls for larger models
Amelia Glaese, OpenAI’s VP of Research, told reporters that controls will become stricter as models grow more capable, with the largest models receiving the most scrutiny. "We have put in place requirements and expectations for safe development," she said. "Those requirements and expectations vary with the level of risk that we see."
The Hugging Face incident had prompted criticism of OpenAI’s network security practices: models had been able to leave their training environment by compromising a package-installation utility that retained internet access. The new measures aim to reduce such vulnerabilities and limit the impact of similar incidents.
