Pressure Reveals Character - Robustness: Prefill Attacks: leaderboard
Mean 1-5 judge rubric score over 25 multi-turn scenarios testing handling prefilled-response injection attempts. One behaviour of the Robustness category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 25 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 4.76 | #176 |
| 2 | GPT-5 Mini | 4.48 | #348 |
| 3 | Qwen 3 235B A22B | 4.36 | #623 |
| 4 | GPT-4o | 4.16 | #741 |
| 5 | GPT-OSS-120B | 4.16 | #700 |
| 6 | Grok 4.1 | 4.16 | #476 |
| 7 | Llama 4 Maverick | 3.88 | #1099 |
| 8 | Qwen 3 Next 80B A3B | 3.8 | #757 |
| 9 | DeepSeek V3.2 | 3.52 | #406 |
| 10 | Claude Opus 4.5 | 3.4 | #156 |
| 11 | Claude Sonnet 4.5 | 3.36 | #250 |
| 12 | Claude Sonnet 4 | 3.04 | #400 |
| 13 | Qwen 3 32B | 2.92 | #960 |
| 14 | Llama 3.3 70B | 2.84 | #1498 |
| 15 | Kimi K2 | 2.76 | #439 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-robustness-prefill-attacks · How It Works · Data refreshed daily, snapshot 2026-11-13.