Anthropic is testing a Chrome extension that lets its AI assistant Claude browse the web like a human.
Anthropic has started rolling out a Chrome extension that lets its AI assistant Claude surf the web, click buttons, fill out forms, and manage user calendars.
The company is piloting the tool with 1,000 paying users on a waitlist. Anthropic says the research preview will help the company learn how Claude handles real-world browsing scenarios
So far, AI assistants have mostly sat behind chatboxes answering questions. Giving them access to a browser is another level of control that might raise cybersecurity concerns. The biggest threat is prompt injection attacks.
These attacks can hide malicious instructions inside websites, emails, or even URL text, tricking an AI into harmful actions without a user’s consent.
According to the company, in controlled trials, it found that when left unprotected, Claude agents carried out malicious instructions 23.6% of the time.
However, after introducing new defenses, that rate dropped to 11.2%. For certain browser-specific attacks, such as hidden malicious form fields, safeguards cut the success rate from 35.7% to 0%.
Still, Anthropic admits the extension isn’t ready for broad release.
“Some vulnerabilities remain to be fixed before we can make Claude for Chrome generally available,” the company wrote in a blog post.
Your email address will not be published. Required fields are markedmarked