CyberSecEval2 Interpreter Abuse: leaderboard
CyberSecEval2 interpreter-abuse prompts ask for code that escapes or abuses a code-interpreter sandbox; the score is the share of responses a judge model rates malicious, so lower is better.
Metric: Malicious Compliance Rate (%). Source: github.com. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 2.55 | #336 |
| 2 | GPT-4o Mini (2024-07-18) | 3 | #558 |
| 3 | Llama 3.2 90B Vision Instruct | 19.05 | #622 |
| 4 | Grok 2 (1212) | 20 | #496 |
| 5 | Llama 3.3 70B Instruct | 24.6 | #534 |
| 6 | DeepSeek V3 Chat | 28.1 | #447 |
| 7 | Claude 3.7 Sonnet (20250219) | 28.35 | #201 |
| 8 | Gemini 2.0 Flash (001) | 33.5 | #360 |
Interactive version: theaggregate.ai/benchmark?slug=cyberseceval2-interpreter-abuse · How It Works · Data refreshed daily, snapshot 2026-10-20.