I'm curious what people on here use to sandbox their coding agents. They all use so many tool calls it's not really possible to check their commands one by one.
I've been using Docker containers but I'm not sure if that's a foolproof solution.
"Gemini" didn't do anything. As we are often admonished in these threads, Gemini is just a stochastic next token predictor. It can't have any responsibility for 'doing' anything.
"A computer can never be held accountable, therefore a computer must never make [an important] decision." -- IBM manual, corrected
Someone hooked up their computer to a random number generator that can destroy it, and their computer got destroyed. The weird part is not why the RNG foot-gun shot someone in the foot again...
I've been using Docker containers but I'm not sure if that's a foolproof solution.
Problem with one website? Change settings for the entire OS.
Problem with one file? Delete the entire folder.
AI is super dangerous when giving advice on OS management. Giving it the keys to your computer is s*icidal.
"A computer can never be held accountable, therefore a computer must never make [an important] decision." -- IBM manual, corrected
Someone hooked up their computer to a random number generator that can destroy it, and their computer got destroyed. The weird part is not why the RNG foot-gun shot someone in the foot again...