Anthropic has announced the launch of a research preview for its new browser-based AI agent, Claude for Chrome, powered by its Claude AI models. This experimental rollout is currently available to a select group of 1,000 users subscribed to Anthropic’s Max plan, which ranges from $100 to $200 per month. For those interested in early access, the company has also opened a waitlist.
Claude for Chrome operates as a sidecar window within the browser, allowing users to interact with the AI while maintaining full context of their browsing activity. Users can optionally grant Claude permission to perform actions on their behalf, such as navigating websites, filling out forms, or completing simple tasks. This marks a significant step toward more agentic AI systems that can actively assist users in real-time.
Anthropic’s move comes amid a growing trend among AI companies to integrate their models directly into browsers. Competitors like Perplexity have launched their own AI-powered browser called Comet, which similarly offers task automation. OpenAI is reportedly close to releasing its own browser with comparable features, while Google has already embedded its Gemini AI into Chrome.
This surge in AI browser development is happening against the backdrop of a major antitrust case against Google, where a federal judge may soon rule that Google must divest its Chrome browser. In anticipation, Perplexity made an unsolicited $34.5 billion bid for Chrome, and OpenAI CEO Sam Altman has expressed interest in acquiring it as well.
Anthropic is approaching this rollout with caution, emphasizing the security implications of giving AI agents browser access. In a recent blog post, the company acknowledged the potential for prompt-injection attacks, where malicious code embedded in websites could manipulate the AI’s behavior. This concern was highlighted last week when Brave’s security team discovered vulnerabilities in Comet’s agent.
To mitigate these risks, Anthropic has implemented several defensive measures. These include:
- Reducing prompt injection success rates from 23.6% to 11.2% through targeted interventions.
- Blocking access by default to sensitive websites, including those related to financial services, adult content, and piracy.
- User-controlled permissions, allowing individuals to restrict Claude’s access to specific sites.
- Requiring explicit user approval before executing high-risk actions such as publishing content, making purchases, or sharing personal data.
This isn’t Anthropic’s first attempt at developing AI agents capable of controlling user interfaces. In October 2024, the company introduced an agent that could operate a PC, though early testing revealed performance issues such as slowness and unreliability.
Since then, agentic AI models have made significant strides. Recent evaluations by TechCrunch suggest that modern agents like Comet and ChatGPT Agent are now fairly reliable for routine tasks, though they still face challenges with complex problem-solving.
