A New Era of Browser-Based AI: Anthropic Launches Claude for Chrome with Enhanced Security Features Anthropic has unveiled a research preview of a browser-based AI agent powered by its Claude models. The company announced the release on Tuesday. Dubbed "Claude for Chrome," this agent is initially available to 1,000 subscribers of Anthropic’s Max plan (ranging from $100 to $200 monthly), with a waitlist open for other interested users. By installing a Chrome extension, selected users can now interact with Claude in a side panel that maintains context across all browser activities. The system allows users to authorize Claude to perform actions automatically or execute tasks on their behalf within the browser interface. Browser integration has become a strategic battleground for AI research labs seeking seamless human-AI interaction. Perplexity recently launched its Comet browser featuring task-automating AI agents. OpenAI is reportedly developing a similar AI-powered browser, while Google has integrated its Gemini AI with Chrome in recent months. The urgency to develop AI-driven browsers intensified following Google’s looming antitrust case, where a federal judge may soon mandate the company to sell Chrome. Notably, Perplexity submitted an informal $34.5 billion acquisition proposal, and OpenAI CEO Sam Altman has stated his organization would consider purchasing the browser. In a Tuesday blog post, Anthropic highlighted emerging security risks associated with browser-accessing AI agents. Last week, Brave’s security team reported vulnerabilities in Comet’s agent, which could be exploited via indirect prompt injection attacks—hidden website code manipulating the agent’s page-processing behavior. Anthropic aims to identify and address such risks through its research preview. Existing countermeasures have already reduced prompt injection attack success rates from 23.6 percent to 11.2 percent. Users can configure settings to restrict Claude’s access to specific websites, with default blocks applied to financial services, adult content, and piracy platforms. The agent also requires user approval before executing high-risk operations like data sharing or purchases. This marks Anthropic’s second foray into screen-control AI agents. In October 2024, the company tested a PC-controlling agent, though early trials revealed performance limitations. Since then, agent capabilities have significantly improved.