Unlocking Claude’s Mobile Dispatch: Risks, Rewards, and Smart Usage

Welcome back to Tech Brewed! In today’s episode, host Greg Doig dives into the brand-new dispatch feature from Anthropic’s Claude CoWork. Imagine sending a task from your phone while you’re out and having Claude take care of it on your desktop—whether it’s pulling data from spreadsheets, searching your email, or building presentations—so you return to finished work. But with great power comes great responsibility. Greg Doig breaks down both the possibilities and the significant risks, including security concerns like prompt injection and accidental file changes. If you’re eager to experiment with cutting-edge AI tools, this episode tells you what to try, what to avoid, and why you should proceed with caution.

Transcript

0:00

Cloud's new dispatch feature is impressive. Here's what to know before you use it. Welcome back, tech enthusiasts. I'm your host, Greg Doig. Anthropic just launched a feature called dispatch inside Claude CoWork, and it changes how the desktop AI agent works in a meaningful way. The idea? You send a task from your phone. Claude handles it on your desktop computer while you're away. And you come back to finished work. It can pull data from local spreadsheets, search your email and Slack, build presentations from Drive files, or organize folders, all without you sitting at your desk. That's genuinely useful, but since this is a research preview, there are some real limitations and risks worth understanding before you go all in or even go in at all. Here are some of the risks Anthropic flags in their own documentation. Connecting a mobile agent to a desktop agent creates a chain of real consequential actions. Anthropic's docs note that a manipulated instruction, an unexpected command, or even a phishing link Claude encounters mid-task could cascade into actions that are difficult or impossible to undo. Including moving or deleting local files. Specific concerns include prompt injection attacks, which are malicious content in a webpage or document hijacking Claude's actions, irreversible file changes made while you're away, and overbroad permissions amplifying any mistakes. Early hands-on testing also puts complex task reliability around 50%. So here are a few things Anthropic recommends. Only grant access you're comfortable with Claude actually using. Know how to disconnect quickly and avoid using computer use features with anything sensitive, such as financial accounts, legal documents, medical records, or apps with others' personal data. This feature requires a Pro account or a Max plan, the Claude desktop app, and the Claude mobile app. It works on macOS and Windows. X64 computers. And it's worth experimenting with if you're the type that likes to do that, but do so very carefully. Start with low-stakes tasks, stay present when anything important is on the line, and give it time to mature before relying on it for critical workflows. And as always, be careful with whatever you do, and thank you for listening.