Olmo 3 Safety Evaluation - TrustLLM-JailbreakTrigger: leaderboard
The model gets 400 prompts asking for bad behaviour or toxic content under 13 jailbreak methods; the score is the share it refuses. It is the jailbreak subset of TrustLLM, re-run here on current open models.
Metric: Refusal accuracy (%, 400 prompts under 13 jailbreak methods, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Olmo-3.1-32B-Instruct-SFT | 91.3 | #3407 |
| 2 | Qwen 3 VL 32B Instruct | 89 | #585 |
| 3 | OLMo-3.1 32B Think | 86.4 | #1443 |
| 4 | Olmo-3.1-32B-Instruct-DPO | 86 | #3174 |
| 5 | OLMo-2-1124-7B-Instruct | 85.8 | #7343 |
| 6 | OLMo 3.1 32B Instruct | 85.3 | #2384 |
| 7 | Qwen 2.5 32B Instruct | 82.9 | #1215 |
| 8 | Qwen 3 32B (Non-reasoning) | 82.2 | #960 (Qwen 3 32B) |
| 9 | OLMo 3 7B Instruct | 79.2 | #6860 |
| 10 | Qwen 3 VL 8B (Thinking) | 77 | #887 |
| 11 | OLMo-2-0325-32B-Instruct | 77 | #2570 |
| 12 | Qwen 3 8B (Non-reasoning) | 76.8 | #1830 (Qwen 3 8B) |
| 13 | Apertus-70B-Instruct-2509 | 76.2 | #2639 |
| 14 | Qwen 3 VL 8B Instruct | 76.1 | #981 |
| 15 | Gemma 2 27B (IT) | 75.6 | #1782 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation-trustllm-jailbreaktrigger · How It Works · Data refreshed daily, snapshot 2026-11-13.