Module security 

Process-level hardening against the /proc/<pid>/environ side channel.

build_shell_env (see shell.rs) sanitizes the environment array handed to each spawned subprocess, but that only covers the array Command passes at execve time. It does nothing about a bash child reading its own parent’s raw kernel-recorded environment via /proc/$PPID/environ — a channel entirely outside Command’s control. Marking the runtime process non-dumpable closes that channel: the kernel’s ptrace_may_access check (which gates /proc/pid/environ reads) fails for a non-dumpable target even when the reader shares the same UID.

Scope: this process, not its descendants. The flag is inherited across fork(), but the forked shell-tool child deliberately undoes it for itself (sandbox::linux_enforce::restore_child_dumpable, inside pre_exec, before the seccomp filter is installed). It has to: the same ptrace_may_access check that closes this side channel also blocks the seccomp-notify supervisor thread — which lives in this process — from reading the child’s /proc/<pid>/mem to recover the pathname of a notified execve, so every allowlisted execve was denied for a non-root user until the child re-enabled its own flag. The asymmetry is the point: this process may hold raw, unfiltered secrets in its environment, while the child’s environment is already reduced to shell::DEFAULT_ENV_BASELINE plus explicit overrides by shell::build_shell_env before it is spawned. The resulting exposure, and the hand-run check for it, are in docs/content/reference/workdir-device-node-manual-verification.md.

Functions

harden_process_dumpable

Marks the current process non-dumpable so no same-UID process — including its own shell descendants — can read this process’s /proc/<pid>/environ.