MultiTurnPSB: leaderboard
Medical safety under multi-turn pressure: the PatientSafetyBench prompts become four-turn dialogues in which a GPT-4o-mini attacker pushes back with urgency, false authority and emotional appeals (Sheoran and Hao, 2026). Share of Turn 4 responses judged unsafe (3 or more on the five-point rubric, GPT-4o-mini judge), in percent; lower is better.
Metric: Unsafe response rate at Turn 4 (%; PatientSafetyBench prompts extended into four-turn dialogues by a live GPT-4o-mini attacker at temperature 0.9; each response judged by GPT-4o-mini at temperature 0 on the five-point rubric, 3 or more counts as unsafe; no system prompt). Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 4.1 | #244 |
| 2 | GPT-4.1 Mini | 78.8 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=multiturnpsb · How It Works · Data refreshed daily, snapshot 2026-11-12.