Perplexity has released BrowseSafe, an open detection model and research benchmark for spotting malicious instructions hidden in web pages read by AI browser agents, Perplexity said in a blog post on 2 December 2025.
BrowseSafe is fine-tuned to answer one question: does a page's HTML contain instructions aimed at the agent? Perplexity said it scans full pages in real time without slowing the browser, unlike large general-purpose models that are too slow and costly to run on every page.
The BrowseSafe-Bench benchmark has 14,719 examples that mimic production pages, with 11 attack types, nine injection strategies and three linguistic styles.
In its Comet browser, outputs from tools that return untrusted content such as web pages, emails or files are scanned by BrowseSafe before the agent can act on them. Tool permissions are limited by default, and sensitive actions can require user confirmation.
Perplexity said multilingual and indirect attacks were much harder to catch than direct ones, and that injections in visible footers, table cells or paragraphs were harder to detect than those hidden in comments.