Anthropic has begun limited testing of a Chrome extension that allows its Claude AI to directly use a web browser. The company said the move is a logical next step after connecting Claude to calendars, documents, and other apps, since so much work takes place inside browsers.
With this new tool, Claude can see webpages, click buttons, and fill out forms. Anthropic is launching the feature as a research preview for 1,000 users on its Max plan, who can join a waitlist for access.
The company stressed that browser-using AI creates security risks, especially from so-called “prompt injection” attacks, where hidden text or instructions on a webpage can trick an AI into harmful actions. In internal red-team tests, early versions of Claude for Chrome carried out malicious instructions, such as deleting emails, when exposed to these attacks. Come on now, AI is smart but it’s not that smart yet.
Anthropic said its new defenses—including stricter permissions, user confirmations for sensitive actions, and blocking of high-risk sites—cut successful attacks from 23.6 per cent to 11.2 per cent. In specialized browser attack scenarios, defenses reduced the rate from 35.7 per cent to zero. Still, the company acknowledged more work is needed before wider release.
For now, Claude for Chrome is being tested only with trusted users in non-critical contexts. Anthropic said feedback from this pilot will guide the development of stronger safeguards and improved classifiers, so browser safety can keep pace as AI capabilities expand.
Interested users can join the waitlist at claude.ai/chrome. Check out the video below to see it in action:
