dfbench: leaderboard
Models look for vulnerabilities in codebases they have not seen; the score is the F1 of their findings (balancing known vulnerabilities found against false reports), from depthfirst's own runs. Unlike CyberGym, which asks for a proof-of-concept input for one known bug, the model has to find the bugs itself.
Metric: Detection F1 on held-out codebases (%). Source: fireworks.ai. Saturation forecast: Not forecast. 10 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | High | 51.5 | #5 |
| 2 | DeepSeek V4.1 Flash | High | 44.5 | #62 |
| 3 | Grok 4.5 | High | 38.9 | #85 |
| 4 | Kimi K3 | Max | 37 | #56 |
| 5 | Grok 4.7 | xHigh | 34 | #47 |
| 6 | MiMo-V2.6-Flash | 33.5 | #155 | |
| 7 | GPT-5.6 Luna | xHigh | 32.2 | #122 |
| 8 | GPT-5.6 Sol | xHigh | 28.3 | #30 |
| 9 | MiMo-V2.6-Pro | 27.2 | #76 | |
| 10 | GPT-6 Sol | xHigh | 24.8 | #26 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #6 GPT-6 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dfbench · How It Works · Data refreshed daily, snapshot 2026-11-16.