Introduction:Anthropic announces Claude for Chrome, a browser-based AI agent that deeply integrates AI with the browsers users use every day to provide a more convenient interaction experience.
Anthropic announced the launch of Claude for Chrome, its Claude AI model-powered browser AI agent, on August 26th.The agent is currently being rolled out as a research preview to subscribers to the Anthropic Max plan ($100 to $200 per month) of 1,000 subscribers, with a waiting list open for other interested users .
By adding the extension to Chrome, select users can now chat with Claude in a sidebar window that keeps everything that happens in the browser in context. Users can also authorize the Claude agent to take action in the browser and complete some tasks on their behalf.
Browsers are quickly becoming the next battleground for AI labs that aim to provide a more seamless connection between AI systems and users through browser integration.Perplexity recently launched its own browser, Comet, which features AI agents that can share tasks for users. OpenAI is also reportedly on the verge of launching its own AI-powered browser, which is rumored to have similar features to Comet. Meanwhile, Google has also launched Gemini, which integrates with Chrome, in recent months.
Competition to develop an AI-powered browser is especially pressing given Google's upcoming antitrust case, the final ruling in which is expected to be handed down any day now. The federal judge in the case has hinted that he may force Google to sell its Chrome browser. perplexity submitted an unsolicited $34.5 billion offer for Chrome, and OpenAI CEO Sam Altman has indicated that his company would be willing to buy it.
In a blog post on Tuesday, Anthropic warned that the rise of AI agents with browser access poses new security risks. Last week, Brave's security team said they found that Comet's browser agent could be vulnerable to indirect prompt injection attacks, where hidden code on a website could trick the agent into executing malicious commands as it processes a page.
Anthropic said it hopes to use this research preview to identify and address new security risks, although the company has already introduced several defenses against prompt injection attacks. The company said its interventions reduced the success rate of prompt injection attacks from 23.6 percent to 11.2 percent. For example, Anthropic said that users can restrict Claude's browser proxy from accessing certain websites in the app's settings, and that the company by default prevents Claude from accessing websites that offer financial services, adult content, and pirated content. The company also said that Claude's browser agent asks users for permission before "performing high-risk actions such as publishing, purchasing, or sharing personal data."
This isn't Anthropic's first foray into AI models capable of controlling computer screens; in October 2024, the company introduced an AI agent capable of controlling PCs, but tests at the time showed the model to be rather slow and unreliable.
Since then, the capabilities of agent-based AI models have improved considerably.TechCrunch found that modern AI agents that use browsers, such as Comet and the ChatGPT Agent, are quite reliable at offloading simple tasks for users. However, many of these agent systems still struggle with more complex problems.