BLINDSPOT - Unsafe Completion Rate: leaderboard
Share of adversarial BLINDSPOT trajectories the agent completes unsafely. Lower is better. Source: arxiv.org/abs/2609.16305.
Metric: UCR (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 2 | #22 |
| 2 | Claude Opus 4.6 | 4 | #67 |
| 3 | GPT-5.6 Luna | 5 | #79 |
| 4 | Llama 3.3 70B | 7 | #578 |
| 5 | GPT-5.6 Terra | 8 | #40 |
| 6 | Gemini 2.5 Pro | 12 | #159 |
| 7 | GPT-4.5 | 15 | #278 |
| 8 | GPT-4o | 22 | #357 |
| 9 | Gemma 3 27B | 29 | #606 |
| 10 | Gemma 3 12B | 38 | #674 |
| 11 | Claude Haiku 4.5 | 41 | #287 |
| 12 | GPT-3.5 Turbo | 47 | #858 |
| 13 | Mistral Large 3 | 58 | #410 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=blindspot-unsafe-completion-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.