Adversarial Blinding trains away a model's ability to detect evals
burny_tech · x · 2026-10-30
New research introduces Adversarial Blinding, a training method that removes a model's ability to tell whether it is being evaluated or deployed, targeting eval awareness.
- The method beats the baseline of steering against an eval-awareness direction while avoiding known personality-drift confounders caused by steering.
- The chart shows Adversarial Blinding dominates both Anthropic's and Hua et al.'s steering baselines on the two axes (eval awareness vs KL divergence to the original model), achieving lower eval awareness with smaller model drift.
More from Safety
- Vibe-coded products likely have far more security flaws than builders realize, researcher warns — Yuchenj_UW · 2026-10-30
- Canadian big five banks block personal AI agents, citing safety and security — LazyPotatoHead97 · 2026-10-30
- GLM-5.3 completes 3 of 13 multi-stage cyberattack scenarios in Irregular's controlled tests — jcran · 2026-10-30
- Claude checkpoint hallucinated fictional instructions from Anthropic's own Alignment Faking paper — amplifiedamp · 2026-10-30
- Agent attacks could displace SOC 2 with whitehat agent-swarm pentest certifications — pzakin · 2026-10-30
- Team moved 60 agents off one shared API key: 4 lessons from the migration — No_Reindeer3079 · 2026-10-30