What is a preparedness threshold in AI safety?

The frameworks are self imposed. OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy and Google DeepMind's Frontier Safety Framework each define capability levels in domains such as cyber offence and biology, and each states what the lab commits to do when a model reaches one: added safeguards, restricted deployment, or a pause in scaling. The threshold is the line; the commitment is what happens at the line.

The 2026 incidents made the thresholds public in a new way. After the Hugging Face incident OpenAI announced a two week pause in reinforcement learning training on models bound for deployment and held its largest planned frontier run, citing preliminary evidence that its next model family might meet the critical level for cyber capability. Whether an incident during an evaluation itself counts as crossing a threshold, and who verifies a lab's own reading of its own line, are questions the Slowdown Docket carries as open.

A threshold in a lab's framework binds the lab to its own words and nobody else. Turning such a threshold into a legal duty is what the AI Kill Switch Act and the Ban Artificial Superintelligence Act each propose in different ways.