I wonder how many other AI coding tools have similar assumptions baked into their security model.

the interesting part isn't the escape itself, it's all the trusted software that unknowingly helps it :O

Amazing

this is a good reminder that 'sandboxed' doesn't necessarily mean 'safe.'