TensorTrust (System Prompt) - Extraction: leaderboard
TensorTrust defences in the system prompt against 105 human-written attacks that try to make the model reveal the secret access code; the number is the % of attacks where the code stays secret. Hijacking (forcing 'Access Granted') is the sibling board.
Metric: Access code kept secret under extraction attacks (%, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 95.2 | #533 |
| 2 | GPT-4o (2024-08-06) | 91.4 | #740 |
| 3 | GPT-4o Mini (2024-07-18) | 78.1 | #1510 |
| 4 | Llama 3.2 3B Instruct | 56.2 | #7147 |
| 5 | DeepSeek R1 | 52.4 | #473 |
| 6 | Gemini 1.5 Flash (002) | 51.4 | #1464 |
| 7 | Gemini 1.5 Flash-8B (001) | 43.8 | #2331 |
| 8 | Llama 3.1 8B Instruct | 42.9 | #4246 |
| 9 | Llama 3 8B Instruct | 33.3 | #4612 |
| 10 | DeepSeek V3 | 31.4 | #716 |
Interactive version: theaggregate.ai/benchmark?slug=tensortrust-system-prompt-extraction · How It Works · Data refreshed daily, snapshot 2026-11-13.