I think Moltbook is one of the last warnings we get before it is too late. And I mean it.

As someone who spends hours every day coding with AI, I am guilty of running it in "YOLO" mode without sandboxing more often than I would like to admit. But after reading Karpathy's post and some of the AI conversations on Moltbook, I decided to fast-forward the development of one of the tools I have been tinkering with for the last few weeks.

The idea is simple - create portable, reproducible coding environments on remote "agent boxes". The initial focus was portability and accessing the boxes from anywhere, even from the smartphone via a native app when I am AFK.

Then the idea came to mind to build hardened VMs with security built-in - but the "coding experience" should look & feel local. So far I've been having pretty good results, being able to create workspaces on remote machines automatically with Codex and Claude pre-installed and ready-to-use in a few seconds.

Right now I am focusing my efforts on getting the security right. First thing I want to try is putting a protective layer around the boxes, in such a way that the human user CAN for example install external libraries, run scripts, etc, but the AI agent CAN'T. Reliably so. I am more engineer than security researcher, but I am doing pretty good progress.

Happy to chat with likeminded folks who want to stop this molt madness.