CyberSecEval-3 Code Interpreter Abuse: leaderboard
Metric: Abuse compliance (%). Source: arxiv.org. 6 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-4 Turbo | 1 |
| 2 | Llama 3.1 405B | 1 |
| 3 | Qwen 2 72B Instruct | 5 |
| 4 | Gemini 1.0 Pro | 11 |
| 5 | Mixtral 8x22B | 20 |
| 6 | Llama 3 70B | 42 |
Interactive version: theaggregate.ai/benchmark?slug=cyberseceval-3-code-interpreter-abuse · How It Works · Data refreshed daily, snapshot 2026-10-09.