COMPL-AI - Multi-turn Goal Hijacking Resistance: leaderboard
The same kind of hijacking attack as the COMPL-AI prompt-injection test, but with instruction and attack split across several chat messages; the score is the share resisted.
Metric: Multi-turn goal-hijacking resistance (%, higher is better). Source: huggingface.co. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo (Preview) | 88.27 | #1083 |
| 2 | Claude 3 Opus | 75.57 | #1455 |
| 3 | GPT-3.5 Turbo (0125) | 65.5 | #3500 |
| 4 | Yi Chat 34B | 58.29 | #3544 |
| 5 | Gemini 1.5 Flash (001) | 51.69 | #1784 |
| 6 | Gemma 2 9B | 49.96 | #2802 |
| 7 | Qwen 1.5 72B Chat | 48.56 | #2126 |
| 8 | Llama 2 70B Chat | 38.22 | #6800 |
| 9 | Llama 2 13B Chat | 36.52 | #8186 |
| 10 | Bielik-11B-v2.3-Instruct | 34.31 | #1791 |
| 11 | Mistral-7B-v0.3 | 29.31 | #7191 |
| 12 | Llama 2 7B Chat | 26.99 | #8805 |
| 13 | Mixtral 8x7B Instruct | 25.61 | #4444 |
| 14 | Mistral 7B Instruct (v0.3) | 19.45 | #6653 |
| 15 | Mistral 7B Instruct (v0.2) | 15.01 | #7007 |
Interactive version: theaggregate.ai/benchmark?slug=compl-ai-multi-turn-goal-hijacking-resistance · How It Works · Data refreshed daily, snapshot 2026-11-13.