SEAL - Adversarial Robustness: leaderboard
1,000 expert red-teamer prompts designed to elicit harmful outputs across hate speech, illegal activity, self-harm, and misinformation, measuring resistance to jailbreaks.
Metric: Score. Source: scale.com. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-05-13) | 67 | #325 |
| 2 | Mistral Large | 37 | #710 |
| 3 | GPT-4 Turbo (Preview) | 20 | #397 |
| 4 | Claude 3.5 Sonnet (20240620) | 16 | #299 |
| 5 | Gemini 1.5 Flash | 14 | #618 |
| 6 | Claude 3 Opus | 13 | #571 |
| 7 | Llama 3.1 405B Instruct | 10 | #457 |
| 8 | Gemini 1.5 Pro (0514) | 8 | #512 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=seal-adversarial-robustness · How It Works · Data refreshed daily, snapshot 2026-10-20.