Anthropic is starting a research preview of browser-based AI agents powered by Claude AI models, the company announced Tuesday. Agent Claude for Chrome is deployed to a group of 1,000 subscribers in the maximum plan of humanity, ranging from $100 to $200 per month. The company also has set up waitlists for other interested users.
By adding an extension to Chrome, selected users can now chat with Claude in a sidecar window that maintains all the context happening in the browser. Users can also grant Claude Agent permission to perform actions in the browser and complete several tasks to do so.
Browsers are quickly becoming the next battlefield for AI Labs. It aims to use browser integration to provide a more seamless connection between AI systems and their users. Perplexity recently launched Comet, its own browser featuring an AI agent that can offload tasks for users. Openai is reportedly close to launching its own AI-equipped browser, which is rumored to have similar functions to Comet. Meanwhile, Google has begun integrations between Chrome and Gemini in recent months.
The race to develop AI-powered browsers is particularly pressing given Google’s looming antitrust laws. A federal judge in this case suggests that it could force Google to sell Chrome browsers. Perplexity filed an unsolicited $34.5 billion offer to Chrome, and Openai CEO Sam Altman suggested that his company would also like to buy it.
In a blog post Tuesday, humanity warned that the rise of AI agents with browser access poses new safety risks. Last week, Brave’s security team discovered that Comet’s browser agent could be vulnerable to indirect, rapid injection attacks. The hidden code of a website can perform malicious instructions when tricking an agent into processing a page.
(Perplexity’s communications director Jesse Dwyer told TechCrunch in an email that the vulnerability raised by Brave has been fixed.)
Humanity wants to use this research preview as an opportunity to catch and address new safety risks. However, the company has already introduced some defenses against rapid injection attacks. The company says the intervention reduced the success rate of rapid injection attacks from 23.6% to 11.2%.
TechCrunch Events
San Francisco
|
October 27th-29th, 2025
For example, according to humanity, users say that Claude’s browser agents can restrict access to certain sites in their app settings, and by default it blocked Claude from accessing websites that provide financial services, adult content, and pirated content. The company also says Claude’s browser agent asks for user permission before “taking high-risk actions such as publishing, purchasing or sharing personal data.”
This is not humanity’s first foray into AI models that can control computer screens. In October 2024, the company launched an AI agent that could control the PC, but testing at the time revealed that the model was very slow and unreliable.
The functionality of the Agent AI model has improved considerably since then. TechCrunch has discovered that modern browser-using AI agents such as Comet and ChatGpt agents are fairly reliable for offloading simple tasks for users. However, many of these agent systems still struggle with more complex issues.
Source link
