Anthropic Updates Responsible Scaling Policy to Version 3.0

Why Anthropic Revised Its Safety Framework

On February 24, 2026, Anthropic released version 3.0 of its Responsible Scaling Policy (RSP), the company’s voluntary framework for managing catastrophic AI risks. The firm says it used two-plus years of operational experience since the 2023 launch to identify what worked, what did not, and what needed to become more explicit. The central theme of the update is practical governance: keep safeguards that proved effective, and increase transparency around decisions made under uncertainty.

From Threshold Logic to Operational Clarity

The original RSP used conditional “if-then” commitments tied to capability thresholds. In practice, this maps to AI Safety Levels (ASLs): if a model crosses a risk threshold, stronger safeguards are required. Anthropic reports that this approach was useful for earlier stages such as ASL-2 and ASL-3. But for higher future levels, the company argues that unilateral implementation can become structurally difficult due to technical uncertainty, ambiguous thresholds, and the need for broader ecosystem coordination.

What Changed in RSP 3.0

• Two-track mitigation model: A clearer split between commitments Anthropic can execute unilaterally and recommendations that require multilateral industry uptake.
• Frontier Safety Roadmap: A published roadmap across Security, Alignment, Safeguards, and Policy, with progress visibility.
• Risk Reports: Systematic reports connecting model capabilities, threat models, and active mitigations, with external expert review in defined cases.

Anthropic also indicates that Risk Reports are intended for public release, with limited redactions only where necessary for legal, privacy, or security reasons.

Why This Matters for the AI Ecosystem

RSP 3.0 is notable because it acknowledges the boundary between what one lab can enforce on its own and what requires policy and industry alignment. That distinction is increasingly important as frontier models become more capable and potential misuse scenarios become harder to mitigate through unilateral controls alone. By adding Frontier Safety Roadmaps and Risk Reports, Anthropic is trying to convert high-level safety principles into recurring, inspectable operating processes.

In short, version 3.0 is less about changing rhetoric and more about building a governance mechanism that can adapt as capabilities advance.

References: Anthropic RSP 3.0 announcement, Responsible Scaling Policy hub