RSP v2 (2023) → v3.0 (Feb 24, 2026): what changed
additions in v3.0
• two-tier mitigation schedule
(Anthropic-unilateral vs industry recommendation)
• AI R&D-4 threshold (research automation)
once crossed: publish affirmative case
• Frontier Safety Roadmap (standing doc)
• Risk Report (standing doc, retrospective)
• RAND SL-4 security referenced
(in the industry-recommendation tier)
• acknowledgement: "confidently ruling out
AI R&D-4 is becoming difficult"
v3.0 announcement notes the threshold is a live concern
even on current-gen Claude Opus 4.6
removals / softenings in v3.0
• 2023 pause commitment REMOVED
(explicit training-pause clause dropped)
• quantitative thresholds softened
to qualitative judgments
• specific mitigations replaced with
"affirmative case" framing
• independent review kept weak
(internal Safety Advisory Group primary)
• industry-recommendation tier allows
policy-advocacy framing without commitment
SaferAI independent rating
v2 (2023): 2.2 — moderate
v3.0 (2026): 1.9 — weak (joins OpenAI Preparedness and DeepMind FSF in "weak" category)
downgrade drivers: qualitative thresholds · pause removal · limited independent oversight
policy gets less rigorous while looking more polished; reading closely is the skill.