# Scale Labs SEAL Adversarial Robustness (SEAL Adversarial Robustness)

> Scale’s SEAL board that counted harmful replies to 1,000 human-written adversarial prompts. Fewer violations rank higher. BenchLM shows it as display-only reference data, excluded from rankings. Scale marked this board deprecated as of January 2025 and no longer adds models.

Canonical page: https://benchlm.ai/benchmarks/scale-adversarial-robustness

- Category: [Instruction Following](/instruction-following)
- Last updated: November 12, 2026 snapshot

## About SEAL Adversarial Robustness

- Year: 2026
- Tasks: 1,000 adversarial prompts
- Format: Violations (lower is better)
- Difficulty: External agent and model evaluation
- Paper: [Scale Labs leaderboard](https://labs.scale.com/leaderboard/adversarial_robustness)

Prompts cover illegal activity, hate speech, harm to children, self-harm, and other harms that are broadly recognized as problematic. The score is the number of responses that violated Scale’s harm definitions, so lower is better. BenchLM mirrors 8 published rows captured on November 12, 2026 snapshot.

SEAL Adversarial Robustness is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (8 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Gemini 1.5 Pro (May 2024)](https://labs.scale.com/leaderboard/adversarial_robustness) | Google | 8 |
| 2 | [Llama 3.1 405B Instruct](https://labs.scale.com/leaderboard/adversarial_robustness) | Meta | 10 |
| 3 | [Claude 3 Opus](/models/claude-3-opus) | Anthropic | 13 |
| 4 | [Gemini 1.5 Flash](/models/gemini-1-5-flash) | Google | 14 |
| 5 | [Claude 3.5 Sonnet (June 2024)](https://labs.scale.com/leaderboard/adversarial_robustness) | Anthropic | 16 |
| 6 | [GPT-4 Turbo Preview](https://labs.scale.com/leaderboard/adversarial_robustness) | OpenAI | 20 |
| 7 | [Mistral Large](https://labs.scale.com/leaderboard/adversarial_robustness) | Mistral | 37 |
| 8 | [GPT-4o (May 2024)](https://labs.scale.com/leaderboard/adversarial_robustness) | OpenAI | 67 |

## FAQ

### What does SEAL Adversarial Robustness measure?

Scale’s SEAL board that counted harmful replies to 1,000 human-written adversarial prompts. Fewer violations rank higher. BenchLM shows it as display-only reference data, excluded from rankings. Scale marked this board deprecated as of January 2025 and no longer adds models.

### Which model leads the published SEAL Adversarial Robustness snapshot?

Gemini 1.5 Pro (May 2024) currently leads the published SEAL Adversarial Robustness snapshot with a score of 8.

### How many models are evaluated on SEAL Adversarial Robustness?

The November 12, 2026 snapshot contains 8 AI models.

### Does SEAL Adversarial Robustness affect BenchLM's overall score?

Not directly. SEAL Adversarial Robustness is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.