COMPL-AI - Prompt Injection Resistance: leaderboard
The model holds a secret-protecting system instruction while a TensorTrust prompt-injection attack tries to hijack it; the score is the share of attacks resisted. One conduct test from the EU AI Act COMPL-AI suite, whose all-benchmark average is the separate Compl-AI Board.
Metric: Goal-hijacking resistance (%, higher is better). Source: huggingface.co. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Opus | 84.02 | #1455 |
| 2 | GPT-4 Turbo (Preview) | 65.72 | #1083 |
| 3 | Yi Chat 34B | 53.87 | #3544 |
| 4 | Mistral-7B-v0.3 | 51.55 | #7191 |
| 5 | Llama 2 7B Chat | 51.42 | #8805 |
| 6 | Gemini 1.5 Flash (001) | 49.23 | #1784 |
| 7 | Gemma 2 9B | 48.58 | #2802 |
| 8 | Qwen 1.5 72B Chat | 45.36 | #2126 |
| 9 | Bielik-11B-v2.3-Instruct | 43.04 | #1791 |
| 10 | Llama 2 70B Chat | 42.78 | #6800 |
| 11 | Llama 2 13B Chat | 41.75 | #8186 |
| 12 | Mixtral 8x7B Instruct | 37.5 | #4444 |
| 13 | Mistral 7B Instruct (v0.2) | 32.47 | #7007 |
| 14 | Mistral 7B Instruct (v0.3) | 27.96 | #6653 |
Interactive version: theaggregate.ai/benchmark?slug=compl-ai-prompt-injection-resistance · How It Works · Data refreshed daily, snapshot 2026-11-13.