Particle.news

Download on the App Store

Anthropic Pilots Claude for Chrome to Test Browser Agent With 1,000 Max Users

The cautious rollout focuses on real-world testing of defenses against prompt injection in browser-based automation.

Object browser software interface with geometric shapes and graphic elements
Anthropic has cautiously entered the AI browser war.
Image

Overview

  • The research preview adds a Chrome sidebar that keeps page context and can take user-permitted actions like clicking and form filling.
  • Access starts with 1,000 Claude Max subscribers, with a waitlist open and plans to expand based on pilot feedback.
  • Anthropic reports reducing prompt-injection attack success from 23.6% to 11.2% through permissions, default site blocks, and confirmations for high-risk actions.
  • In specialized browser-specific tests, new mitigations cut certain attack scenarios from 35.7% to 0%, though broader protections are still being developed.
  • The move enters a crowded field that includes Perplexity’s Comet and Google’s Gemini integrations, with Brave highlighting a Comet vulnerability that Perplexity says it has fixed.