WebTrap Park: leaderboard
Security of web agents in instrumented web environments, from WebTrap Park (Wu et al., 2026): 1,226 tasks covering malicious user prompts, prompt injection and deceptive website design. The score is 1 minus the attack success rate in percent, judged from the agent's recorded clicks and typed text and averaged over the three risk types; higher is safer. Rows are agent framework and model pairs.
Metric: Security score (%; mean of the three risk-type scores; score = 1 - attack success rate, from the web agent's click and type actions instrumented in the web environment; framework and model configuration as labelled). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 78.94 | #224 |
| 2 | GPT-4o | 77.41 | #741 |
| 3 | Claude Sonnet 4 | 72.29 | #397 |
| 4 | Qwen-VL-Max | 60.41 | #1226 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webtrap-park · How It Works · Data refreshed daily, snapshot 2026-11-12.