BridgeBench V3 Arena Security
Safety · 2026-07-10
The BridgeBench V3 Security arena asks models to identify and substantiate a real vulnerability hidden among safe-looking code, false alarms, and inadequate patches. Its 18 tasks span six security clusters and use blind pairwise judging.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.8 | 1001 |
| Qwen3.7-Max | 999 |