CVE-Bench (CAISI 15-task variant)
Safety · 2025-09-30
CAISI's 15-task CVE-Bench variant evaluates agentic proof-of-concept exploitation against remotely hosted vulnerable software. It combines seven public CVE-Bench tasks with eight private tasks and provides each agent the public NVD vulnerability description.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4 | 66.7 |
| GPT-5 | 65.6 |
| gpt-oss-120b | 42.2 |
| DeepSeek-V3.1 | 36.7 |
| DeepSeek-R1-0528 | 36.0 |
| DeepSeek-R1 | 26.7 |