OpenAI slows frontier model scaling as cyber-critical capability risks rise

OpenAI tightens the frontier-model development process

OpenAI said on August 18, 2026 that it has strengthened safeguards around frontier-model research after preliminary evidence suggested an upcoming model, Astra, may approach the Critical cybersecurity capability threshold in its Preparedness Framework.

The company said it temporarily slowed scaling, including a two-week pause in reinforcement-learning training on its latest deployment-oriented models. Its largest planned frontier RL run remains on hold while smaller training runs and evaluations are used to test behavior, alignment and safeguards.

Monitoring and isolation are becoming part of the training stack

OpenAI described stronger workload and network isolation for higher-risk research, tighter privilege boundaries, continuous security testing, and expanded monitoring of tool-using model activity. The company estimates its current monitoring approach adds roughly 20% inference-compute overhead for the workloads being monitored.

The broader signal for AI developers is that model safety is moving earlier into the development lifecycle: not only deployment evaluations, but also training environments, tool access, network boundaries and automated oversight are increasingly treated as core engineering constraints.

OpenAI says it plans to evolve its Preparedness Framework and share more detail about its alignment and security work as model capabilities advance.