Katarzyna Petru
Claude for Chrome – test version for selected users
Anthropic is launching a new AI agent that operates directly in Google Chrome. Claude for Chrome has reached 1,000 subscribers of the Max plan ($100–$200 per month). The rest of the interested parties can sign up for the waiting list.
The extension allows for conversations with Claude in a side window that retains the context of the browsed pages. Additionally, the agent – with the user's consent – can perform certain tasks in the browser.
Race for the AI Browser
The browser is becoming a new battlefield. Perplexity already has its own Comet, Google has added Gemini integrations, and OpenAI is preparing a similar solution. The stakes are high, especially since Google may be forced to sell Chrome as a result of antitrust proceedings. Perplexity has already made a purchase offer of $34.5 billion, and Sam Altman from OpenAI has also suggested interest in an acquisition.
Security and Prompt Injection Attacks
Anthropic acknowledges that agents with browser access carry serious risks. An example is the recent vulnerabilities in Comet from Perplexity, which could have allowed websites to trigger unwanted actions from the agent.
The company has already implemented safeguards – the effectiveness of prompt injection attacks has decreased from 23.6% to 11.2%. Additionally, Claude has default access blocked to financial sites, adult content, or pirated material. When engaging in "high-risk" activities, such as making purchases or sharing personal data, the agent always asks for consent.
Another Step After the Agent on PC
This is not Anthropic's first attempt in this direction. In 2024, the company showcased an agent that could control a computer, but it was slow and not very stable. Today, the situation looks different – agents operating in the browser are becoming increasingly useful. Comet and ChatGPT Agent handle simple tasks, although complex challenges are still a problem.
