Anthropic is taking the next step in AI development by enabling Claude to work directly inside your browser. After months of integrating Claude with calendars, documents, and apps, the company has now launched a controlled pilot for a Chrome extension that enables Claude to perform tasks such as clicking buttons, filling out forms, and managing online workflows.

Why Browser-Using AI Matters?

Most modern work happens in browsers. By allowing Claude to interact directly with the browser, users can streamline tasks such as scheduling meetings, drafting emails, or handling reports. However, this innovation raises significant safety concerns, particularly regarding malicious attacks designed to deceive AI systems into performing harmful actions.

Tackling Safety Challenges

One of the most significant risks is prompt injection attacks, which involve hidden instructions embedded in websites, emails, or documents that can mislead AI into deleting files, stealing data, or making unauthorized transactions.

In testing, Anthropic found a 23.6% attack success rate when Claude was targeted without safeguards in place. However, with new safety mitigations, the rate dropped to 11.2%, and for browser-specific attacks, some scenarios achieved a 0% success rate.

Defenses include

• Permissions & Confirmations: Users control which sites Claude can access and must confirm high-risk actions.

• Restricted Sites: Claude is blocked from high-risk categories, including financial and adult content.

• Advanced Classifiers: Detect unusual patterns in instructions to stop suspicious activity.

How Users Can Participate?

The project is currently in limited testing with 1,000 Max plan users. Trusted testers can join the Claude for Chrome waitlist and, once approved, install the extension via the Chrome Web Store.

Participants are encouraged to use Claude on trusted sites only, avoid sensitive environments such as financial or medical platforms, and share their feedback. Insights will help refine defenses, improve safety classifiers, and shape the future of browser-integrated AI.

The Road Ahead

Browser-using AI is an inevitable evolution of AI assistants. By addressing safety challenges early, Anthropic aims to build trustworthy, secure, and practical tools that can handle real-world browsing tasks without compromising security.

This research preview is not just about testing features; it’s about laying the foundation for safer AI-driven browsing that benefits both individuals and businesses.