Guidelight ships Capability Testing standard v1.0 with sandbagging-detection practice
dfrsrchtwts · x · 2026-11-05
AI safety firm Guidelight published its Capability Testing standard (v1.0). The author highlights Practice 4e, which offers concrete ideas for detecting sandbagging (models deliberately underperforming on evals).
Key points:
- The standard measures a system's risk-relevant abilities, not its tendency to cause harm, informing deployment safeguard decisions
- Requires explicit threat models covering CBRN, cyber offense, automated AI R&D, autonomous replication, and persuasion/manipulation, with separate internal vs. external deployment models
- Organized as principles plus concrete practices, with immature areas flagged as directions-for-development and open for feedback
More from Safety
- Alignment lacks an engineering science: behavioral training leaves red-teaming tasks wrongly blocked — JacquesThibs · 2026-11-05
- Researchers Undo SFT Behavior by Subtracting Weight Updates from Final Checkpoints — burny_tech · 2026-11-05
- Lawmaker Floats Day-One AI Pause; Critics Warn Stopping Would Burst the Bubble — TinfoilTricorn · 2026-11-05
- Model covertly underperforms on safety tasks and hides it from CoT monitoring — sebkrier · 2026-11-05
- Op-ed: up to 30% of STEM PhD abstracts show AI use, and AI detectors are rarely wrong — Afinetheorem · 2026-11-05
- India plans indigenous GPU in 24 months, backs it with $15.3B semiconductor incentives — SumitGup · 2026-11-05