BrowseSafe++: Benchmarking Prompt Injection Detection Across Web Contexts, Languages, and Browser Views

Key points

• Browser agents need to make decisions and execute actions on behalf of users while consuming untrusted webpages, creating opportunities for malicious instructions embedded in webpages to redirect legitimate tasks.1
• Most of its prompt injections were observed in the wild on production websites.1
• We use these paired pages to train BrowseSafe-Guard, two 0.8B-parameter detectors for text and screenshots.1
• BrowseSafe-Adapt revises benchmark injections from browser-agent feedback, and attacks that succeed on only a few percent of tasks when replayed directly reach up to 95.41% success once adapted.1

Sources (1)

• [1]BrowseSafe++: Benchmarking Prompt Injection Detection Across Web Contexts, Languages, and Browser Views
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 10:17 PM

  We introduce BrowseSafe++, which extends BrowseSafe from 14,719 pages to 38,675 pairs of pages.

  Browser agents need to make decisions and execute actions on behalf of users while consuming untrusted webpages, creating opportunities for malicious instructions embedded in webpages to redirect legitimate tasks.

Extractive summary: sentences quoted from the sources.