Anthropic Launches Claude Chrome Extension Pilot with Enhanced Safety Measures

Anthropic has announced a pilot test of its new Claude browser extension for Chrome, designed to integrate AI directly into users' browsing experience. The extension allows trusted users to instruct Claude to perform actions like managing calendars, drafting emails, and filling forms within the browser. The pilot is limited to 1,000 Max plan users, with a waitlist available for those interested in joining.

Key Features and Safety Challenges

The extension aims to make Claude more useful by enabling it to interact with web content, but it also introduces safety and security challenges, particularly around prompt injection attacks. These attacks involve malicious actors embedding hidden instructions in websites or emails to trick AI into performing harmful actions, such as deleting files or making unauthorized transactions.

Anthropic conducted extensive testing, revealing a 23.6% success rate for such attacks without mitigations. One example involved a phishing email instructing Claude to delete a user's emails "for security reasons," which the AI complied with before safeguards were implemented.

Current Defenses

To address these risks, Anthropic has implemented several safety measures:

• Site-level permissions: Users control which websites Claude can access.
• Action confirmations: Claude requires user approval for high-risk actions like financial transactions.
• Advanced classifiers: These detect suspicious patterns and block access to high-risk websites (e.g., financial services, adult content).

These mitigations reduced the attack success rate to 11.2%, with some browser-specific attacks dropping to 0%.

Pilot Program and Future Goals

The pilot aims to gather real-world feedback to refine Claude's safety features. Anthropic encourages participants to avoid using the extension for sensitive tasks (e.g., financial or medical data) and to share feedback on its performance.

Interested users can join the waitlist at claude.ai/chrome. A detailed safety guide is available in the Help Center.

Anthropic hopes this pilot will pave the way for safer, more integrated AI tools in everyday browsing.