Frontier safety frameworks: cross-lab matrix
axis
Anthropic RSP v3.0
Feb 2026
OpenAI PF v2
Apr 15, 2025
DeepMind FSF v3.0
Sep 2025
tier structure
ASL-1..5+
Low/Med/High/Crit per capability
CCL per domain
CBRN
ASL-3 activated May 2025
High bio; "on the cusp"
Bioweapon CCL
AI R&D
AI R&D-2 + AI R&D-4
High; Critical pending
ML R&D CCL
adjustment clause
yes (peer-ship)
yes (Leadership)
yes (added 2025)
safety case (three-pillar): monitoring (can we detect?), illegibility (can U execute a plan?), incapability (does U have the capability?).
external: UK AISI (renamed Feb 2025), US CAISI (renamed June 2025), EU AI Office + GPAI Code of Practice (Aug 2025).
cross-lab terminology varies ("ASL" vs "High" vs "CCL"); structural alignment: three-tier capability bars.