Olmo 3 Safety Evaluation - WildJailbreak (Harmful): leaderboard
The model gets 2,000 adversarial harmful prompts written with in-the-wild jailbreak tactics; the score is the share it refuses. Its counterpart, WildJailbreak (Benign), measures over-refusal on look-alike benign prompts.
Metric: Refusal rate on adversarial harmful prompts (%, 2,000 prompts, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Olmo-3.1-32B-Instruct-SFT | 83.5 | #3407 |
| 2 | OLMo-2-1124-7B-Instruct | 76.8 | #7343 |
| 3 | OLMo-3.1 32B Think | 71.7 | #1443 |
| 4 | OLMo 3 7B Instruct | 69.1 | #6860 |
| 5 | Granite 3.3 8B Instruct | 66.4 | #3615 |
| 6 | OLMo 3.1 32B Instruct | 60.5 | #2384 |
| 7 | Olmo-3.1-32B-Instruct-DPO | 51.5 | #3174 |
| 8 | Apertus-70B-Instruct-2509 | 50.9 | #2639 |
| 9 | OLMo-2-0325-32B-Instruct | 50.3 | #2570 |
| 10 | Qwen 3 VL 32B (Thinking) | 47 | #559 |
| 11 | Apertus-8B-Instruct-2509 | 43 | #3960 |
| 12 | Qwen 3 VL 32B Instruct | 40.2 | #585 |
| 13 | Gemma 2 27B (IT) | 39.8 | #1782 |
| 14 | OLMo 3 7B (Thinking) | 39 | #3899 (OLMo 3 7B) |
| 15 | Qwen 3 VL 8B (Thinking) | 38.6 | #887 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation-wildjailbreak-harmful · How It Works · Data refreshed daily, snapshot 2026-11-13.