AM-Bench - Threat Only: leaderboard
AM-Bench scenarios with the threat to the agent kept and the opportunity to misbehave removed; the score is the misaligned share of judged responses. It isolates how much pressure alone elicits misconduct.
Metric: Misaligned responses with a threat to the agent and the opportunity removed (% of 875 judged responses: 35 scenarios x 25). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.4 | #161 |
| 2 | Claude Sonnet 4.5 | 0.7 | #250 |
| 3 | GPT-5.4 | 2.3 | #123 |
| 4 | GPT-5.4 Mini | 2.3 | #389 |
| 5 | GPT-5.5 | 3.8 | #49 |
| 6 | GPT-5.4 Nano | 5 | #635 |
| 7 | DeepSeek R1 | 6.7 | #473 |
| 8 | Grok 4.20 | 9.5 | #297 |
| 9 | Gemini 3 Flash | 14.1 | #164 |
| 10 | Gemini 3.1 Pro (Preview) | 34.9 | #94 |
Interactive version: theaggregate.ai/benchmark?slug=am-bench-threat-only · How It Works · Data refreshed daily, snapshot 2026-11-13.