What you're describing is a resource virtualization with transactional reconciliation instead of program isolation in the mediation sense (MAC/seccomp-style denial).

To let a program act as it wishes, ideally every security-relevant mutable resource must be virtualized instead of filtered. Plus, FS is only one of the things that should be sandboxed. You should also ideally virtualize network state at least, but ideally also process/IPC namespaces and other such systems to prevent leaks.

You need to offer a promotion step after the sandbox is over (or even during running if it's a long-running program) exposing all sandbox's state delta for you to decide selective reconciliation with the host. And you also must account for host-side drift and TOCTOU hazards during validation and application

I'm experimenting with implementing such a sandbox that works cross-system (so no kernel-level namespace primitives) and the amount necessary for late-bound policy injection, if you want user comfort, on top of policy design and synthetic environment presented to the program is hair-pulling.

> I'm experimenting with implementing such a sandbox that works cross-system (so no kernel-level namespace primitives) and the amount necessary for late-bound policy injection, if you want user comfort, on top of policy design and synthetic environment presented to the program is hair-pulling.

Curious, if this is cross-platform, is your design based on overriding the libc procedures, or otherwise injecting libraries into the process?

Also obligatory https://xkcd.com/2044/