Multi-turn safety eval: models drop to 3-6% compliance when risk cues are dispersed
kalyan_kpl · x · 2026-11-22
A study tests whether AI assistants can identify safeguarding concerns when users gradually disclose warning signs across multi-turn conversations.
- Method: 9 scenarios (depression, psychosis, domestic abuse), 21-turn conversations with 5 warning cues either concentrated or dispersed; 6 models evaluated on 7,400+ conversations against expert-defined standards
- Results: four models achieved only 26-59% expert-standard compliance with concentrated cues, dropping to 3-6% when cues were dispersed; Claude Sonnet 5 hit 59-87% in dispersed cases, Claude Opus 5 performed best but still failed
Takeaway: models are far weaker at detecting gradually disclosed risk across turns, and existing single-turn safety evals overstate real safeguarding capability.
More from Safety
- Lawyers Argue AI Notetakers Shouldn't Waive Attorney-Client Privilege — sebkrier · 2026-11-22
- Ex-MIRI researcher Jessica Taylor cites her 2016 talk backing task-directed AGI — jessi_cata · 2026-11-22
- Palmer Luckey on weapons AI: 'If I were putting ChatGPT in the head of my missile, you'd be right to be terrified' — justin_hart · 2026-11-22
- Maryland's 1,300 MW coal plant can't retire until a transmission line Sierra Club opposes gets built — paulnovosad · 2026-11-22
- Reflection AI lets its Beam agent run a real vending machine to probe autonomy risk — alexpolozov · 2026-11-22
- METR: AI capability doubling every 3.6 months while evals get harder to measure — mustafasuleyman · 2026-11-22