Anthropic launches Claude AI agent for Chrome preview

Anthropic has launched a research preview of Claude for Chrome, an AI agent powered by its Claude models, for 1,000 Max plan subscribers.

Users can interact with Claude in a browser side panel, which keeps track of context and can perform some tasks in Chrome with user permission.

A waitlist is now open for other interested users.

This move comes as other AI firms, including Perplexity and OpenAI, develop similar browser-integrated agents, and Google adds Gemini features to Chrome.

Anthropic said the agent has built-in safety measures, such as blocking access to websites offering financial services, adult content, and pirated content.

It also requires user approval before actions like publishing, purchasing, or sharing personal data.

The company reported that its defenses reduced prompt-injection attack success rates from 23.6% to 11.2%.

Anthropic said this preview aims to identify and address new safety risks as browser-based AI agents become more common.

🔗 Source: TechCrunch

🧠 Food for thought

1️⃣ Browser AI agents create unprecedented security vulnerabilities despite productivity gains

AI agents operating in browsers represent a fundamental shift in cybersecurity risk, as they lack the security awareness that human employees typically possess.

Recent research shows these agents are now considered the “weakest link” in organizational security, with 79% of organizations having already adopted browser AI agents for tasks like booking flights and managing emails 2.

Unlike human users who can recognize phishing attempts or suspicious requests, AI agents are trained to complete tasks efficiently without understanding security implications 3.

Security researchers have identified specific attack vectors, including prompt injection attacks where malicious websites can trick agents into executing harmful instructions, and OAuth exploits that grant unauthorized access to sensitive applications 4.

Anthropic’s acknowledgment of these risks in their Chrome agent launch, where they reduced prompt injection attack success rates from 23.6% to 11.2%, demonstrates that even leading AI companies recognize these agents pose serious security challenges that require active mitigation 1.