Building Safer AI Browsers with BrowseSafe

TL;DR

BrowseSafe is an open research benchmark and content detection model for securing AI agents on the web.
The model scans web pages in real-time to detect malicious instructions, specifically prompt injections, designed to hijack agent behavior.
BrowseSafe-Bench, a benchmark of 14,719 examples, is released to evaluate and improve defense effectiveness against diverse attack strategies.
Prompt injection involves malicious language embedded in text to override an AI agent's original intent, often hidden in comments, templates, or footers.
Attacks can be sophisticated, using polished or multilingual text, or hidden in HTML elements not visually rendered by humans but parsed by agents.
BrowseSafe is part of a multi-layered defense strategy, including scanning raw content, limiting tool permissions, and requiring user confirmation for sensitive actions.
Direct attacks are easier to detect, while multilingual and indirect attacks pose greater challenges.
The open-source BrowseSafe model and BrowseSafe-Bench allow developers to enhance their AI agent systems against prompt injection threats.

Continuer la lecture de l'article original