Agent-style AI can run shell commands. You have to accept them but some people live dangerously and say Yes To All.

I've been letting Gemini run gcloud and "accept all"ing while I've been setting some things up for a personal project. Even with some limits in place it is nervewracking, but so far no issues and it means I can go and get a cup of tea rather than keep pressing OK. Pretty easy to see how easy it would be for rogue AI to do things when it can already provision its own infrastructure.

Sadly, this was the last time anybody heard from PickledChris.

"Open the brine valve HAL."

"I'm sorry Chris. I'm afraid I can't pickle that."

Yep, it's not as far fetched as it would've been a year ago. A scenario where you're running an agent in 'yolo mode', it opening up some poisonous readme / docs / paper, and then executing the wrong shell command.

Could be done responsibly if you run it in a VM to sandbox it with incremental backup so you can roll-back if something is deleted?

If you cheat using an "agent" using an "MCP server", it's still rm -rf on the host, but in a form that AI startups will sell to you.

MCPs are generally a little smarter than exposing all data on the system to the service they're using, but you can tell the chatbot to work around those kinds of limitations.

Do you know that most MCP servers are Open Source and can be run locally?

It's also trivial to code them. Literally a Python function + some boilerplate.

I was sort of surprised to see MCP become a buzz word because we’ve been building these kinds of systems with duck tape and chewing gum for ages. Standardization is nice though. My advice is just ask your LLM nicely, and you should be safe :)