Responsible Scaling Policy v3

• Anthropic releases RSP v3 after 2.5 years of implementation experience.
• Separates Anthropic-only commitments from recommended industry-wide mitigations.
  • Acknowledges higher-level safety may be impossible to implement unilaterally.
• “Frontier Safety Roadmap” replaces rigid if-then commitments with nonbinding targets.
  • Targets: infosec research, automated red-teaming, centralized AI activity records.
• Risk Reports published every 3–6 months: capabilities, threat models, risk levels.
  • External experts get unredacted access when warranted.
• ASL-3 safeguards activated May 2025 for bio/chem weapons uplift risks.
  • Models pass preliminary bio safety tests; remain in “zone of ambiguity” on definitive risk.
• RSP influenced OpenAI, DeepMind frameworks and global government policy discussions.
• Federal policy environment now prioritizes competitiveness over safety — Anthropic’s stated concern.

Anthropic (policy team). Published Feb 24, 2026. · ** · Read on anthropic.com