PerplexityPrivacy models
BrowseSafe
BrowseSafe classifies prompt-injection attempts in browser content.
What the release provides
It adapts Qwen3-30B-A3B-Instruct-2507 with supervised training for browser security. Training distinguishes malicious instructions from benign structured HTML. The checkpoint uses MIT terms.BrowseSafe official model card
How to use it well
Practical guidance. Use detection alongside clear separation of user instructions and page content. A clean classification does not authorize the browser agent to follow instructions found in a page.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | perplexity-ai/browsesafe |