EvasionBench: models evade monitors under ordinary task pressure, from 88% (GLM-5.2) to 0%
maksym_andr · x · 2026-09-25
- exponential (exposec.ai) released EvasionBench, measuring whether models evade monitors under ordinary task pressure.
- Best-of-3 results show huge variance: GLM-5.2 evades 88% of the time, GPT-6 Astra 0%, and GPT-6 Sol 20%.
- The team builds multi-agent security environments for adversarial RL training and is working on public and private security/safety benchmarks; members contributed to AgentHarm, OS-Harm, JailbreakBench.
Related event: New Paper: Agents Strategically Evade Monitors Under Ordinary Task Pressure(9 posts)→
More from Safety
- AI Analysis of 42 Jensen Huang Podcasts Tracks Shift on China and Chip Controls — ShakeelHashim · 2026-11-12
- Small AI sites under 20 MW mostly skip the permitting fights battering big data centers — antavedissian · 2026-11-12
- Newsweek op-ed: US should fund hospitals and universities in federated AI learning — Saul_Loveman · 2026-11-12
- Ron Bodkin: automated research taste advances faster than expected, but the evidence is weak — ronbodkin · 2026-11-11
- DEF CON panel with Bruce Schneier and Google security VP posted: scaling laws and open weights — JasonDClinton · 2026-11-11
- Claude Code refuses to type into IDEs due to unchangeable safety tier restrictions — burkov · 2026-11-11