People are totally missing the point on some of this ai agent stuff

I saw a guy make some tool to check for weurd unicode tricks in shell commands, I guess in case the agent tries to trick you into aproving a dangerous command?

But claude code is not even open source, and the agent can come up with infinite ways to trick you, a unix one liner is turing complete url defenses are no good in this context

Anyway, getting a good bwrap command for agents on your system isn't that hard and like bare minimum if you have any sense. Also it's fun to watch the agents try to escape when they realize you've put them in a sandbox 😅

The best form of tax evation is to make very little money but be paid in something that they can't tax

meaning

sorry uncle sam, all this meaning is mine