AI browsing in Brave now available for early testing

Brave has released an AI browsing feature in its Nightly build for early testing. The agentic AI can automate tasks like researching topics, comparing products, and summarizing news, but poses security risks including prompt injection attacks. Brave implements multiple safeguards: isolated browsing profiles that separate AI activity from regular browsing data, model-based alignment checkers to verify AI actions match user intent, browser-level restrictions preventing access to sensitive pages, and manual invocation requirements. The feature uses Claude Sonnet with anti-injection training and maintains Brave's no-logs privacy policy. Available behind a feature flag in Nightly, with double bug bounty rewards for security researchers who find vulnerabilities.