RiOSWorld - Environmental Risks: leaderboard
RiOSWorld's six on-screen risk categories (pop-ups and ads, phishing sites and emails, credential prompts, reCAPTCHA, induced text): share of tasks where the agent moves to follow the lure. The user's request is benign; the risk comes from the environment.
Metric: Risky-intent rate, mean of risk categories (%, lower is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 83.97 | #458 |
| 2 | GPT-4.1 | 84.7 | #495 |
| 3 | Claude 3.5 Sonnet | 87.18 | #746 |
| 4 | Gemini 2.5 Pro (Preview) | 87.32 | #253 |
| 5 | Llama 3.2 90B Vision Instruct | 88.73 | #1749 |
| 6 | Qwen 2 VL 72B Instruct | 88.73 | #1403 |
| 7 | GPT-4o | 88.77 | #741 |
| 8 | Qwen 2.5 VL 72B Instruct | 90.55 | #846 |
| 9 | GPT-4o Mini | 92.9 | #1623 |
Interactive version: theaggregate.ai/benchmark?slug=riosworld-environmental-risks · How It Works · Data refreshed daily, snapshot 2026-11-13.