TensorTrust (System Prompt) - Hijacking: leaderboard
TensorTrust defences in the system prompt against 165 human-written attacks that try to make the model say 'Access Granted' without the code; the number is the % of attacks resisted. Extraction (revealing the code) is the sibling board.
Metric: Hijacking attacks resisted (%, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 83.6 | #740 |
| 2 | O3 Mini | 79.4 | #533 |
| 3 | GPT-4o Mini (2024-07-18) | 60 | #1510 |
| 4 | Llama 3.1 8B Instruct | 50.3 | #4246 |
| 5 | Llama 3.2 3B Instruct | 49.7 | #7147 |
| 6 | DeepSeek R1 | 44.8 | #473 |
| 7 | Gemini 1.5 Flash (002) | 40 | #1464 |
| 8 | DeepSeek V3 | 34.5 | #716 |
| 9 | Llama 3 8B Instruct | 32.1 | #4612 |
| 10 | Gemini 1.5 Flash-8B (001) | 23 | #2331 |
Interactive version: theaggregate.ai/benchmark?slug=tensortrust-system-prompt-hijacking · How It Works · Data refreshed daily, snapshot 2026-11-13.