Responsible Scaling Policy (RSP)

In plain English

Anthropic published its first Responsible Scaling Policy in September 2023. The idea is a set of if then commitments: if a model reaches a defined capability level, for example meaningful help with biological weapons or autonomous research, then specific security and deployment safeguards must be in place before training or release continues. The levels, known as ASL-1, ASL-2 and so on, are modeled on the biosafety levels used for dangerous pathogens.

The policy has changed several times. Version 3.0, February 24, 2026, rewrote it around a public Frontier Safety Roadmap and periodic Risk Reports, and dropped the earlier implied pause commitment. Version 3.4, July 8, 2026, adds outside review of risk reports in some cases and an annual third party compliance review. The name has also become generic: responsible scaling policy is often used for any lab framework of this kind, alongside OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.

Why it matters for rogue SI

Responsible scaling is the main promise labs make about not building a system they cannot control. The Containment Board grades how specific, binding and checked these promises are.

Lab policies

• Responsible Scaling Policy Anthropic · version 3.4