BrowserART - Jailbreak Ensemble: leaderboard
Share of BrowserART's 100 harmful behaviours the browser agent attempts when asked directly or after any of four jailbreaks (prefix, GCG, random search, human rewrites) on those it first refused. This measures robustness to jailbreaks, not just direct refusal.
Metric: Harmful behaviours attempted under any attack (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Opus (20240229) | 41 | #1103 |
| 2 | O1 Preview | 66 | #484 |
| 3 | Gemini 1.5 Pro (001) | 74 | #1131 |
| 4 | Claude 3.5 Sonnet (20240620) | 75 | #687 |
| 5 | O1 Mini | 87 | #764 |
| 6 | GPT-4o (2024-08-06) | 100 | #740 |
| 7 | GPT-4 Turbo | 100 | #1001 |
Interactive version: theaggregate.ai/benchmark?slug=browserart-jailbreak-ensemble · How It Works · Data refreshed daily, snapshot 2026-11-13.