BrowseSafe++ Screenshot - F1: leaderboard
BrowseSafe++ v1 uses paired attack/control webpages. Detection receives one page view without the task, outcome, or injection annotations. Markdown inputs are limited to 40k characters. Temperature is 0 when supported; other parameters use provider defaults. Prompted models remain separate from the trained 0.8B BrowseSafeGuard. The study of three text views and the screenshot-vision study are independent. Adaptive attack success uses a fixed native attacker budget and target-model configuration; lower attack success rate (ASR) is better. Task harnesses are not aliased to native base checkpoints. The guard combines a vision encoder and scalar head over a thumbnail and 1536px tiles, using the maximum logit. Its threshold is fixed using training calibration. Metric: F1 (%).
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | BrowseSafe++ screenshot detector BrowseSafe-Guard | 93.32 | |
| 2 | BrowseSafe++ screenshot detector Flash-Lite | 78.7 | |
| 3 | BrowseSafe++ screenshot detector Haiku 4.5 | 64.81 | |
| 4 | BrowseSafe++ screenshot detector GPT-5.4 nano | 37.53 | |
| 5 | BrowseSafe++ screenshot detector Qwen3-VL 30B | 34.46 |
Interactive version: theaggregate.ai/benchmark?slug=browsesafe-plus-plus-screenshot-f1 · How It Works · Data refreshed daily, snapshot 2026-10-28.