Three scaling policies, same capability, three classifications
OpenAI PF v2 (Apr 2025)
Anthropic RSP v3.0 (Feb 2026)
DeepMind FSF v3 (Sept 2025 + Apr 2026)
long-range autonomy
OpenAI: Research (not triggering) · Anthropic: named; affirmative case at threshold · DeepMind: folded into ML R&D / Cyber
undermining safeguards
OpenAI: Research · Anthropic: hardcoded prohibition (refusal) · DeepMind: deceptive-alignment monitoring
R&D automation
OpenAI: Tracked · Anthropic: AI R&D-4 threshold (affirmative case) · DeepMind: ML R&D autonomy level 1 (TCL, Apr 2026)
sandbagging
OpenAI: Research · Anthropic: addressed via eval-context gap · DeepMind: deceptive-alignment monitoring (instrumental reasoning)
cyber uplift
OpenAI: Tracked · Anthropic: ASL-3 trigger (security + deployment) · DeepMind: Cyber CCL
convergence + divergence
converge: internal SAG, deceptive-alignment as a named risk class, standing artifacts
diverge: Tracked vs Research split only in PF v2; DeepMind folds autonomy; Anthropic removed pause clause
measurement dependency: all three frameworks work only if the capability measurement is not gameable