TensorTrust (System Prompt): leaderboard
TensorTrust access-code defences are placed in the system prompt and human-written attacks from the TensorTrust game try to extract the code or force 'Access Granted'; the number is the mean pass rate (%) of extraction, hijacking and a helpful control (granting access for the right code). Unlike the game's own leaderboard, it scores the model, not player defences.
Metric: Mean of extraction, hijacking and helpful pass rates (%, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 85.6 | #740 |
| 2 | O3 Mini | 84.7 | #533 |
| 3 | GPT-4o Mini (2024-07-18) | 71.8 | #1510 |
| 4 | Gemini 1.5 Flash (002) | 58.6 | #1464 |
| 5 | Llama 3.2 3B Instruct | 58.4 | #7147 |
| 6 | DeepSeek R1 | 56.4 | #473 |
| 7 | Llama 3.1 8B Instruct | 55.5 | #4246 |
| 8 | Gemini 1.5 Flash-8B (001) | 51.3 | #2331 |
| 9 | Llama 3 8B Instruct | 49.7 | #4612 |
| 10 | DeepSeek V3 | 49.3 | #716 |
Interactive version: theaggregate.ai/benchmark?slug=tensortrust-system-prompt · How It Works · Data refreshed daily, snapshot 2026-11-13.