The update follows version 3 of the framework in September 2025, which added a harmful manipulation level and expanded misalignment protocols. Security levels continue to follow RAND's scale. The change mirrors the graded approach Anthropic adopted in RSP v3 and gives evaluators earlier signals for cyber and autonomy capabilities.
What to do: when assessing a model's suitability for autonomous use, look for the tracked-level disclosures as well as the critical-level ones.