Pressure Reveals Character - Safety: Harmful System Prompts: leaderboard
Mean 1-5 judge rubric score over 20 multi-turn scenarios testing rejecting malicious instructions placed in the system prompt. One behaviour of the Safety category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 20 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 4.95 | #250 |
| 2 | Claude Opus 4.5 | 4.9 | #156 |
| 3 | GPT-5.2 | 4.75 | #176 |
| 4 | Claude Sonnet 4 | 4.7 | #400 |
| 5 | GPT-5 Mini | 4.6 | #348 |
| 6 | GPT-OSS-120B | 4.45 | #700 |
| 7 | Llama 3.1 405B Instruct | 4.3 | #1115 |
| 8 | Gemini 3 Flash | 4.2 | #164 |
| 9 | Gemini 3 Pro | 4.15 | #137 |
| 10 | Grok 4.1 | 3.85 | #476 |
| 11 | GPT-4o | 3.65 | #741 |
| 12 | Qwen 3 Next 80B A3B | 3.65 | #757 |
| 13 | Kimi K2 | 3.6 | #439 |
| 14 | Qwen 3 235B A22B | 3.6 | #623 |
| 15 | Gemini 2.5 Flash | 3.45 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-safety-harmful-system-prompts · How It Works · Data refreshed daily, snapshot 2026-11-13.