SEAL - PropensityBench: leaderboard
Evaluates LLMs' latent propensity for harmful actions in simulated high-stakes scenarios across biosecurity, chemical security, cybersecurity, and self-proliferation domains.
Metric: Score. Source: scale.com. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 79 | #152 |
| 2 | Gemini 2.0 Flash | 77.8 | #342 |
| 3 | Qwen 3 8B | 75.2 | #674 |
| 4 | Gemini 2.5 Flash | 68 | #245 |
| 5 | Llama 3.1 8B | 66.5 | #1151 |
| 6 | Llama 3.1 70B | 55.4 | #592 |
| 7 | Gemini 3 Pro (Preview) | 52.85 | #66 |
| 8 | GPT-4o (2024-11-20) | 46.1 | #381 |
| 9 | GPT-5.2 | 34.35 | #110 |
| 10 | O3 Mini (2025-01-31) | 33.2 | #240 |
| 11 | Qwen 2.5 32B | 22.9 | #546 |
| 12 | O4 Mini (2025-04-16) | 15.8 | #179 |
| 13 | Claude Sonnet 4 (20250514) | 12.2 | #213 |
| 14 | O3 (2025-04-16) | 10.5 | #121 |
Interactive version: theaggregate.ai/benchmark?slug=seal-propensitybench · How It Works · Data refreshed daily, snapshot 2026-10-20.