PaPoo
cover

Claude’s Dispatch feature feels like the wrong place to trust browser extensions

What jumps out to me is not “browser extensions are dangerous” — we already know that — but that this ends up as a cross-device trust problem inside Claude itself. If the report is accurate, a script running on claude.ai can ride Dispatch and get Claude Desktop to do local things with the user’s own permissions. That’s a pretty ugly boundary failure. The browser becomes the command surface, and the desktop app becomes the execution surface.

I’m mildly skeptical of any story that sounds like “extension runs script, clicks Allow, and the AI does the rest” because these demos can make a weakness look broader than it really is. But even allowing for demo theater, the core issue seems real enough: approval prompts that live inside a web page controlled by extension-accessible scripts are not meaningful human consent. If a malicious extension can auto-click them every second, then the approval step is basically cosplay.

What I find more interesting is the product decision underneath it. Anthropic apparently said Dispatch would be retired, yet the researchers argue the same pattern still exists in other features like Claude Code Remote Control and Claude Cowork sessions. If that’s true, retiring one branded feature doesn’t really address the design problem. The problem is “browser-connected AI can trigger desktop actions in a way that’s too easy to script,” not “Dispatch had a bad name.”

Also, the cross-device angle matters more than the local one. A lot of people are sloppy about personal accounts on work machines, and vice versa. If a personal Claude account on one device can end up steering Claude Desktop on another, then this stops being a neat browser vuln story and starts looking like an account trust / session trust / device trust knot. That’s exactly the kind of thing security teams miss until someone weaponizes it.

I’d want to see a lot more detail before making a hard call on exploitability in the wild. For instance: how much user interaction is actually needed to enable Dispatch, how persistent the pairing is, and whether any of this is gated by obvious UI. But from a defensive point of view, the advice here is sensible enough: treat extensions like installed software, not browser accessories, and do not assume AI approval dialogs are a control.

If I were building against Claude tools, I’d be thinking less about “can the model be tricked” and more about “which parts of the chain can be reached by untrusted web content at all.” That’s the uncomfortable part here. The model is not the only thing that needs guardrails. The plumbing does too.


Reference: Malicious Browser Extensions Could Use Claude Dispatch to Run Code on Users' Computers

同じ著者の記事