⚠️ Important: AI browsing is currently an experimental feature available only in Brave Nightly for early testing. Agentic web browsing has inherent security and privacy risks that we discuss in our series on “Security & Privacy in Agentic Browsing”. Use with caution and avoid sensitive tasks.

AI browsing gives your browser autonomous capabilities where it can perform tasks on your behalf. Instead of just answering questions or summarizing content, AI browsing can actively browse websites, research information across multiple sources, fill shopping carts, compare products, and complete multi-step tasks—all while you maintain control over its actions.

As a privacy-first company, Brave has implemented multiple safeguards to protect users while testing this transformational technology. AI browsing operates in an isolated browser profile, uses alignment checking to verify intended actions, and maintains Brave's commitment to not logging or retaining your data.

What is AI Browsing?

AI Browsing enables Leo to act autonomously within the browser to accomplish tasks you specify. Rather than simply providing information, AI Browsing can:

• Research topics by visiting and analyzing multiple websites
• Compare products across different shopping sites
• Load items into shopping carts
• Check for valid promo codes before purchases
• Gather information from various sources to create comprehensive reports
• Perform fact-checking by investigating claims across the web
• Complete multi-step workflows that would normally require manual browsing

AI browsing works seamlessly with Brave's existing features, including Skills. For example, you could create a /fact-check Skill that instructs AI to assess an article by researching supporting evidence, or a /check-promo-codes Skill that searches for discount codes before you complete a purchase.

Requirements

• Platform: Desktop only (Windows, macOS, Linux)
• Browser version: Brave Nightly 1.87 or higher
• Subscription: No Leo Premium subscription required for testing
• Future availability: Android and iOS support planned for future releases

Important Security & Privacy Information

Understanding the risks:

Agentic browsing is inherently risky. Giving an AI agent control of your browsing experience could potentially expose personal data or allow the agent to take unintended actions. While Brave has implemented multiple security measures, these safeguards do not eliminate all risks, including prompt injection attacks where malicious actors embed instructions in web content.

The two primary risks are:

1. Malicious actors: Attackers who attempt prompt injection on websites to manipulate the agent's behavior
2. Model confusion: The AI model misinterpreting commands and taking unintended actions

Brave has chosen a careful, staged approach to releasing AI browsing, making it available first in Nightly (our testing build) behind an opt-in flag to gather feedback from early testers and security researchers.

How Brave protects you:

Brave implements multiple layers of security to mitigate these risks:

Isolated storage: AI browsing operates in a completely separate browser profile with its own cookies, logged-in state, caches, and site data. This isolation prevents the agent from accessing your regular browsing profile, protecting your existing logged-in sessions and personal data even if other defenses fail.

Alignment checking: A second AI model reviews the agent's planned actions before execution, comparing them against your original request to ensure the agent is following your intentions. This alignment checker is designed to catch potential prompt injections and unintended behaviors, though it cannot eliminate all risks.

User control: AI browsing must be manually invoked and cannot activate on its own. The interface clearly shows when the agent is active and what it's doing with options for manual intervention in the conversation view. Actions flagged as potentially risky by the alignment checker trigger warnings and require your permission before proceeding.

Access restrictions: AI browsing does not have access to your main browser profile, internal browser pages (such as brave://settings), non-HTTPS pages or websites flagged by Safe Browsing.

Data privacy: AI browsing follows Brave's standard privacy protections. Conversations are not logged, retained, or used for model training. Your IP address is hidden through a reverse proxy, and no personal data is stored on Brave's servers. Support for locally-hosted models or other 3rd party APIs via BYOM will be implemented in a future release.

How do I enable AI browsing?

AI browsing is available in Brave Nightly behind a feature flag. Follow these steps to enable it:

1. Download and install Brave Nightly (version 1.86 or higher)
   
   
2. Open Brave Nightly and enter brave://flags in the address bar
3. Search for #brave-ai-chat-agent-profile in the flags search box
4. Click the dropdown next to Brave AI browsing and select Enabled
   
   
5. Click the Relaunch button to restart your browser and apply the changes

Once enabled, you'll see a new AI browsing icon in the Leo sidebar.

How do I use AI browsing?

Starting an AI browsing session:

1. Open Brave AI by clicking the AI icon in the sidebar or by going to brave://leo-ai
2. Click the AI browsing icon to launch AI browsing
   
   
3. A new browser window will open with the isolated AI browsing profile
   
   
4. AI browsing is now active and ready to assist you

Using AI browsing:

Once AI browsing is active, you can interact with it just like regular AI, but with autonomous capabilities:

• Direct requests: Ask Brave AI to complete tasks that require browsing. 例如：
  • "Research the top three CRM platforms and create a comparison report"
  • "Find the best price for [product name] across major retailers"
  • "Check if there are any working promo codes for [store name]"
• With Skills: Combine AI browsing with your saved Skills for powerful automation:
  • Type /fact-check to have the agent verify claims in an article by researching sources
  • Type /check-promo-codes before checkout to search for discount codes
  • Type /competitive to analyze multiple competitor websites and generate insights
• Multi-step tasks: AI browsing can handle complex workflows:
  • "Add [product 1], [product 2], and [product 3] to my cart on [store name]"
  • "Research recent developments in [topic] from the past month and summarize key findings"

AI browsing works in both sidebar and full-page views. All agent actions occur visibly in the chat window so you can see exactly what the agent is doing.

Managing your session:

• Inspect actions: Watch AI browsing progress in real-time through the chat interface
• Pause or stop: You can interrupt the agent at any time if it's not doing what you intended
• Consent prompts: For sensitive actions that might have been injected by an attacker, AI browsing will ask for your explicit permission before proceeding
• Session data: The isolated profile and its data persist after you close the AI browsing window appearing as a selectable browser profile for future use

To leave your AI browsing session and return to regular browsing, click the "AI Profile" icon on the top-right of the window and select the profile you want to return to:

Common use cases

For researchers and students:

• Multi-source verification: "Check these statistics against three different reputable sources and note any discrepancies"
• Comprehensive research: "Research recent academic papers on [topic] and create a summary of current thinking"
• Fact-checking: Create a /fact-check Skill to automatically verify claims by researching supporting evidence

For online shoppers:

• Price comparison: "Find [product] on Amazon, Best Buy, and Walmart and compare prices including shipping"
• Promo code hunting: Create a /check-promo-codes Skill to search for valid discount codes before purchase
• Product research: "Compare the top 5 noise-cancelling headphones under $300 based on recent reviews"

For professionals:

• Competitive analysis: Use with the /competitive Skill to research competitor offerings and create detailed comparisons
• Market research: "Research the top 10 players in [industry] and summarize their key differentiators"
  Report compilation: "Gather information about [topic] from industry news sites and create an executive summary"

Advanced combinations

Combine AI browsing with Leo's other features for enhanced capabilities:

• AI browsing + Skills: Create custom Skills that leverage AI browsing capabilities (e.g., /research-brief, /competitive, /fact-check)
• AI browsing + Multi Tab Context: Start with multiple tabs open, then ask AI browsing to analyze and synthesize information across all of them
• AI browsing + Memories: Enable Leo's memory features so AI browsing can reference your preferences and past research in future tasks

Known limitations

• Rate limits: Complex tasks may require many requests. If you hit a rate limit during a task, simply instruct the agent to "continue" or "keep going" to resume the task.
• Model constraints: AI browsing capabilities are limited by the underlying AI model's abilities. It may not always interpret complex requests correctly on the first try.
• Access restrictions: AI browsing cannot access internal browser pages (brave://settings, brave://flags, etc.), non-HTTPS websites and pages tagged by Safe Browsing, for security reasons.
• Experimental status: As an early testing feature, AI browsing may exhibit unexpected behaviors. This is why it's currently limited to Nightly with multiple safeguards in place.

Troubleshooting

AI browsing icon not appearing:

• Ensure you've enabled the #brave-ai-chat-agent-profile flag in brave://flags
• Verify you've restarted Brave Nightly after enabling the flag
• Check that you're using Brave Nightly version 1.86 or higher

Agent gets stuck or stops responding:

• Check if you've hit a rate limit—if so, instruct the agent to continue
• Try rephrasing your request more specifically
• If the agent seems confused, start a new session with a clearer prompt

Agent asks for permission repeatedly:

• This is expected behavior for sensitive actions
• Review what the agent is trying to do before granting permission
• If unexpected, decline permission and rephrase your request

Concerned about agent behavior:

• Remember you can pause or stop the agent at any time
• Review the chat Steps to see what actions were taken
• If you encounter a security issue, report it to Brave's bug bounty program

Managing AI browsing data

Profile persistence: The isolated AI browsing profile remains available as a selectable browser profile even after you close the AI browsing window. You can return to it later or clear its data through standard profile management.

Clearing data: To remove data from your AI browsing profile:

1. Go to Brave Settings
2. Navigate to the profile management section
3. Select the AI browsing profile
4. Choose to clear browsing data or remove the profile entirely

Session logs: AI browsing maintains logs of your sessions within the isolated profile. You control this data and can delete it at any time.

Help us improve AI browsing

AI browsing is in active development, and we value your feedback:

Report security issues: If you discover a security vulnerability in AI browsing, please report it to our bug bounty program. During this early release phase, valid and in-scope security issues in AI browsing receive double our usual reward amounts.

Share feedback and bugs: As an open-source browser, we welcome bug reports and feature requests on GitHub.

Learn more: Read our series on Security & Privacy in Agentic Browsing to understand our approach to building safe agentic experiences.

Related Documentation

For more information about Leo's features and capabilities:

• 我如何使用 Brave Leo？
• How do I use Skills in Brave Leo?
• How to Use Multi Tab Context with Brave Leo
• Leo 的 AI 模型之間的差異是什麼？
• Security & Privacy in Agentic Browsing (blog series)