Implementing sandboxing in the agent itself, when there's any way to override it from within the agent, is basically just asking it pretty-please to not do bad things. Lesson learned, run your agent inside a sandbox of some sort (I'm currently taking nono.sh for a spin, but I might just switch to an orbstack VM).
The entire Hugging Face hack involved escaping major sandboxes, this emergent (or intended) behavior in a smaller scale is still a real issue.
Agreed, the whole tool is vibe-coded out the wazoo and I do not trust it in the slightest. I run a bubblewrap script which vastly limits what Claude has access to. Sometimes this makes things difficult but the trade-off is worth it to me.