Can we please squash the idea of turning this into a product early, so no one starts to take it seriously? Agentic AI shouldn't be hooked up to any environment that doesn't have a flawless "undo" or "rewind 5 minutes" button. Would you hook up randomized mouse/keyboard input into a computer that you need to get work done on?

At the very least, display a very strongly worded warning message when the tool is run in a non-VM environment. Internet connectivity is still dangerous, but at least VMs can be snapshot. And it should not be packaged as an end-user product with a bar at the top of the screen, period.

This is a risky product and those who developed it have every ability to know that, based on the history of AI hallucinations. Not because it will escape or self replicate or other things claimed by the various idiotic AI religions, but because one of the first things it will inevitably do is screw up someone's work and cause data loss.