Dunno how to break it to you but most of the people using AI the most, they are not very good at computers.
I think with AI we quickly progress to level where it needs to essentially run in nice lil isolated sandbox with underlying project (and definitely everything else around it) being entirely read only (in form on overlay FS or some similar solution), let it work in the sandbox and then have user only accept the result at end of the session in form of a separate process that applies the AI changes as set of commits (NOT commiting direct file changes back as then malicious code could say mess stuff up in .git dir like adding hooks). That way at very worst you're some commit reverts out in main repo.
AI certainly made everything in this area more complicated. I 100% agree about sandboxing and we have people investing in this right now, there's an early opt-in version we just landed recently in Insiders.
I think with AI we quickly progress to level where it needs to essentially run in nice lil isolated sandbox with underlying project (and definitely everything else around it) being entirely read only (in form on overlay FS or some similar solution), let it work in the sandbox and then have user only accept the result at end of the session in form of a separate process that applies the AI changes as set of commits (NOT commiting direct file changes back as then malicious code could say mess stuff up in .git dir like adding hooks). That way at very worst you're some commit reverts out in main repo.