Responsible Scaling Policy Version 3.0

Anthropic has released the third version of its Responsible Scaling Policy, the voluntary framework it uses to reduce catastrophic risks from its AI systems. The company says it has had such a policy for over two years and, after assessing its benefits and shortcomings, is keeping what has worked while improving the rest and adding measures meant to make its decision-making more transparent and accountable.

In reviewing the original design, which rested on if-then commitments tied to AI Safety Levels, Anthropic says some goals were met. The policy drove stronger safeguards, such as classifiers deployed under its ASL-3 standard, which it activated for relevant models in May 2025. within months OpenAI and Google DeepMind took up roughly comparable frameworks, and laws including California's SB 53 and New York's RAISE Act, along with codes of practice under the EU AI Act, now call for published risk frameworks. Other goals fell short. Capability thresholds proved ambiguous, with biology offering a zone where risk can be neither confidently ruled out nor established. Government action on safety has been slow, and higher safeguard levels may be impossible for one company to achieve alone, as a RAND report on model weight security suggests.

The new version has three main elements. First, it separates mitigations Anthropic will pursue regardless of others from a broader capabilities-to-mitigations map intended for the whole industry. Second, it requires a Frontier Safety Roadmap covering security, alignment, safeguards and policy, framed as public, nonbinding goals whose progress will be graded openly. Third, it introduces Risk Reports, published with some redactions every three to six months, with external expert review required in certain circumstances; pilots of such review are already running.

Anthropic describes the policy as a living document that it will keep revising as the technology develops.