The permission model only helps in production when scopes match what the process actually needs, not a blanket allow. Prefer --allow-fs limited to the directories the service reads and writes, keep --allow-net pinned to the hosts your outbound clients call, and fail closed on unexpected module paths so a dependency cannot open sockets or files you never intended.
Ops prove-out after tightening scopes: boot the same binary under the new flags and run a canary that exercises disk and outbound calls you expect to pass, then confirm a deliberate out-of-scope path is denied before any side effect. Watch crash and restart rates for one deploy window — a spike usually means a transitive import still opens a temp path or metrics endpoint outside the allow list. Roll under peak and compare successful request rate; if health probes fail while app paths pass, the probe binary needs its own scoped allow list rather than inheriting a wider one.