Anthropic has launched a trial version of a new browser agent powered by its Claude AI model. The "Claude for Chrome" agent is currently available to 1,000 subscribers of Anthropic's "Max" plan, which costs $100-200 per month. The company has also created a waitlist for other interested users.
This extension allows users to chat with Claude in a side window that monitors their browser activity. Users can also grant the agent permission to perform certain actions and automatically complete tasks on their behalf.
Browsers have become the new competitive arena for AI companies. These firms are aiming to create easier connections between their AI systems and users through browser integrations.
Perplexity recently launched its own AI browser, "Comet," designed to simplify user tasks.
Google has added Gemini integrations to Chrome in recent months.
This competition has intensified, especially due to Google's ongoing antitrust case. The court’s upcoming decision could force Google to sell its Chrome browser. Perplexity has even made an unsolicited $34.5 billion offer for Chrome, and OpenAI CEO Sam Altman has indicated his company would also be willing to buy it.
Safety Concerns
Anthropic warns that AI agents with browser access pose new security risks. For example, Brave's security team found a vulnerability in the Comet browser agent. This flaw could have allowed hidden code on a website to force the agent to perform malicious actions. Perplexity has stated that this issue has since been resolved.
Anthropic plans to use this trial version to discover and address new safety risks. They have already implemented safeguards that reduced the success rate of prompt injection attacks from 23.6% to 11.2%. Users can also limit Claude’s access to specific sites like those offering financial services, adult content, or pirated material. Furthermore, Claude will always ask for user permission before taking high-risk actions such as publishing, making purchases, or sharing personal data.
This is not the first time Anthropic has developed an AI model that can control a computer screen. They introduced a similar agent in October 2024, but it was found to be very slow and unreliable during testing. Since then, the capabilities of agentic AI models have significantly improved. TechCrunch's tests show that modern browser agents like Comet and ChatGPT Agent are quite reliable for simple tasks, but they still struggle with more complex problems.
