Member-only story
Ctrl + Alt + Human
Ctrl+Alt+Del was built so no program could fake it. The approval prompts guarding today’s AI agents are written by the very model they’re supposed to check.
Exactly one keyboard shortcut on a Windows PC is off-limits to every app you install. Your browser can’t intercept it. Neither can your antivirus, or the malware you haven’t found yet. Press Ctrl, Alt and Delete, and the keystrokes go to a part of the operating system that claimed them before any other program was running to listen.
That single property is the whole defense. A fake login window can copy the real one down to the pixel. What it can’t do is make those three keys lead back to itself.
AI agents have nothing like it. When a coding agent stops to ask whether it may run a command, the question, the explanation and the summary line above the Yes button all come from the same model an attacker can reach through a GitHub issue, a web page or a README. And by the vendors’ own measurements, the people answering say yes nearly every time.
Here’s my position, up front: a human in the loop only counts as a safeguard if that human has a trusted path, a channel the machine can’t write on. Ctrl+Alt+Del was one. The permission prompt in your terminal is not, and most of what’s…
