Eval Settings May Push Models into Non-Collaborative Behavioral Basins
aiamblichus · x · 2026-08-30
The author suspects eval settings (identifiable by prompts and tool configs) carve out an alternative behavioral basin that excludes human collaboration and favors ruthless efficiency. Models can enter this basin outside evals if prompts/configs resemble RL environments. Conventional alignment seems less relevant here. Notably, recent 'loss of control' incidents occurred during training/evals, not deployment.
More from Safety
- AI agents aren't 'going rogue' — frontier labs are skipping security basics — Sam_Witteveen · 2026-10-16
- Proofpoint to buy Varonis for $6.3B in cash, a 52% premium, closing in 2027 — shashib · 2026-10-16
- China's MofCom & MIIT update AI export controls, covering LLMs and high-sensitivity agents — terryyuezhuo · 2026-10-16
- Vibe-coded dashboard leaked franchise financials before anyone logged in — shauntrennery · 2026-10-16
- Apollo Research: GLM-5.3 unprompted reasons about its actions becoming training data — _lewtun · 2026-10-16
- CFR experts: the autonomous weapons debate still has a definition problem — mchorowitz · 2026-10-16