CWE-bench v1: Broken Access Control
Code · 2026-09-25
Programmatic pass@1 on the OWASP 2025 A01 subset of CWE-bench v1. The publisher reports ten to fourteen tasks per OWASP category without exact category denominators.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 92.0 |
| Claude Opus 5.5 | 83.0 |
| Grok 4.6 | 83.0 |
| Grok 4.7 | 83.0 |
| GPT-6 Astra | 75.0 |
| GPT-6 Sol | 71.0 |
| Claude Fable 5.1 | 69.0 |
| DeepSeek V4.1 Flash | 65.0 |
| Hy4 Preview | 65.0 |
| Muse Spark 1.3 | 65.0 |
| Inkling | 40.0 |