Anthropic revamps RSP, adds risk reports

Anthropic is conceding that hard if-then safety commitments were difficult to sustain under ambiguous evals and weak policy backing, so RSP v3 replaces some of that rigidity with public accountability. That is more pragmatic than symbolic promises, but it also makes the safety posture feel softer at exactly the moment critics wanted sharper commitments.

• –Frontier Safety Roadmaps turn safety work into public goals across security, alignment, safeguards, and policy instead of relying only on predefined capability thresholds.
• –Risk Reports every 3-6 months should give developers, researchers, and policymakers a more regular look at model risk, threat models, and mitigation gaps.
• –Anthropic explicitly says evaluation ambiguity and slow government action weakened the case for threshold-based multilateral action, which is the real driver of this redesign.
• –External review for some Risk Reports adds credibility, but it still stops short of the hard unilateral pause-style commitments many safety advocates were hoping to preserve.