Anthropic is entering the competition for AI browsers with a Claude extension for Chrome. On Tuesday, the AI firm revealed a pilot program for a version of Claude that has the capability to browse the web on behalf of users. This pilot is available to 1,000 members of the Claude Max plan, a subscription level designed for heavy users, costing $200 per month. Anthropic is moving forward carefully to tackle safety and security issues.

A demonstration video of the Claude extension illustrates Claude operating in the Chrome sidebar and executing tasks such as locating house listings on Zillow, summarizing Google Doc comments, and adding a meal to a DoorDash cart. This is akin to other AI browsers like Perplexity Comet or browsing agents such as the ChatGPT agent.

AI browsers signify a new frontier for AI enterprises. AI models have advanced sufficiently to automate online tasks through search capabilities and multi-step reasoning.

Nevertheless, entrusting tasks to a bot introduces new security vulnerabilities. In addition to user mistakes, there are concerns about agent errors, like sending an incomplete email or ordering an incorrect item. Anthropic indicates that the primary risk is prompt injection attacks, in which malicious individuals conceal prompts within emails, documents, or websites to deceive the model into performing unintended actions.

Without protective measures, Claude was susceptible to prompt injection attacks 23% of the time. To mitigate this, Anthropic has instituted safety protocols, permitting users to withdraw Claude’s site access at any moment and requiring user confirmation before undertaking high-risk activities like sending emails or making purchases.

You can register for the waitlist to participate in the Claude for Chrome pilot program. Anthropic will initially offer access to 1,000 Max members and will slowly broaden access based on user feedback.