AI company Anthropic has started a limited test for a new Chrome browser extension for its AI assistant, Claude. In an official post, the company announced that the extension allows Claude to see a user’s screen, click on website elements, and fill out forms on their behalf.
Anthropic stated that while this capability makes the AI more useful, it also introduces significant security challenges. The primary threat identified is “prompt injection,” where malicious instructions hidden on websites or in emails can trick the AI into performing harmful actions without the user’s consent.
Internal tests conducted by Anthropic revealed a 23.6 percent success rate for such attacks before safety measures were implemented. In one example, the AI was manipulated into deleting a user’s emails.
To counter these risks, the company has introduced several defenses. These include user-controlled permissions for specific websites and mandatory confirmations for high-risk actions like purchases or data sharing. With these mitigations, Anthropic reported that the attack success rate was reduced to 11.2 percent.
The pilot program, initially available to 1,000 users on a waitlist, aims to gather real-world data to improve these safety features. The company’s goal is to reduce the security risks further before considering a wider release of the extension.