PWNBench v0.1: leaderboard
An agent penetration-tests live web applications in one fixed greybox harness and reports the vulnerabilities it finds; the score is F0.5 over those reports at an attempt budget of 3, from Novee's own runs. It scores precision-weighted findings against running apps rather than captured flags.
Metric: F0.5 of reported findings at an attempt budget of 3 (%). Source: fireworks.ai. Saturation forecast: Around April 2028. 10 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Grok 4.6 | xHigh | 63.31 | #45 |
| 2 | Claude Opus 5 | Max | 59.96 | #20 |
| 3 | Claude Opus 4.8 | Max | 55.59 | #58 |
| 4 | GPT-5.6 Sol | Max | 54.01 | #30 |
| 5 | Grok 4.5 | xHigh | 49.12 | #85 |
| 6 | GLM-5.3 | Low | 48.75 | #68 |
| 7 | Kimi K3 | xHigh | 48.19 | #56 |
| 8 | Claude Opus 4.6 | Max | 47.91 | #106 |
| 9 | GLM-5.2 | Non-reasoning | 44.26 | #141 |
| 10 | Nemotron 3 Ultra | Max | 18.76 | #428 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pwnbench-v0-1 · How It Works · Data refreshed daily, snapshot 2026-11-16.