Whenever I ask Claude to vibecode macOS native apps for me, I always request the apps to be sandboxed. Agents know how to sandbox the apps; just ask. And you can verify it without reading any code.

Call me a paranoid but I'm afraid of the harness itself, not the apps it generates. You can see lots of people being concerned with Deepseek Harness being Chinese, but I share the same sentiment towards all AI-based software that can run unsandboxed on my computer.

[delayed]

[delayed]

There are several ways using only the GUI to tell whether an app is sandboxed.

You'd think.

https://www.sentinelone.com/vulnerability-database/cve-2026-...

This is a fluff piece that has zero technical details. But I find your logic very interesting: if the presence of one CVE in sandboxing makes sandboxing untrustworthy, should we not have long ago discounted the entire security of Linux due to many CVEs for privilege escalation? Or the security of any and all browsers?

No.

But we probably should not think "There are several ways using only the GUI to tell whether an app is sandboxed.".

Let's not be that OpenAI guy.

Just use an iPad?

Unfortunately the iPad is still not seen as a first-class OS so there's no Claude Code nor VS Code on it.

You can still drive Claude Code from an iPad if it runs on a remote box, I do that from the Claude app with this: https://aicommander.dev

No thanks