BrowseSafe: Understanding and Preventing Prompt Injection Within AI Browser Agents