AI companies don't bother securing their super scary product
I'm being made to use Claude for my work and one thing that has struck me is how bad the sandboxing is for a tool they are worried about.
The permissions model is beyond basic, for bash it's just string matching that doesn't even support regexes, but a lot of my work involves calling APIs which would also be relatively easy* to properly sandbox (filtering on API paths and verbs and regexes on payloads, etc) and it also doesn't offer that.
*Beyond my skill level but you could do it for less than $30B
Basically if they believed their own bullshit, they would be doing a far better job.













