fleet/fleetd
3
0
Fork 0
Code Issues 73 Pull Requests Actions Packages Projects Releases 1 Wiki Activity

scrub: stop export UID= aborting the allow-list scrub mid-loop #396

Closed
agent wants to merge 1 commits from fix/scrub-uid-abort into main
pull from: fix/scrub-uid-abort
merge into: fleet:main
Conversation 2 Commits 1 Files Changed 2
+63 -2

1 Commits

Author SHA1 Message Date
ltms d678783af7 scrub: stop export UID= aborting the allow-list scrub mid-loop
CI / contract (pull_request) Successful in 1m23s
Details
CI / build (pull_request) Successful in 1m32s
Details 
CB-633's generated scrub.zsh blanks every non-allow-listed exported name
in one loop wrapped in `{ ... } 2>/dev/null`. In zsh, `export UID=` is
not a failed command: it is a fatal parameter error ("failed to change
user ID: operation not permitted") that aborts the whole sourced file.
The loop stops at UID, every later name is left unscrubbed, and the
report-writing block never runs -- silently, because of the 2>/dev/null.

The inversion is the severity. `env` lists inherited names first and the
names a startup file exports last, so the loop blanked the harmless
inherited half and died immediately before the operator's own exports --
exactly the credentials this policy exists to remove.

Measured on a fleet01 member pane: UID is name 42 of 57, and a decoy
exported from ~/.zshrc sat at 58, non-empty, on 8 of 8 spawns. xtrace
ends at `+scrub.zsh:28> export 'UID='` with rc=126. Reproduced
independently on macOS/zsh 5.9 (fleetd #394).

CONTAINMENT: `eval`, then VERIFY.

Neither `export "$n=" 2>/dev/null || true` nor a `${(t)n}` type guard
contains it -- both still abort, because it is an assignment error and
not a command failure. Only `eval` does. `eval` is safe at that point
because the loop above already rejected every name that is not
[A-Za-z_][A-Za-z0-9_]*, so nothing but a bare identifier reaches it.

Enumerating the specials (UID|EUID|GID|EGID|PPID|LINENO) also works, but
only for the names enumerated: one that turns up exported on another
host brings the abort straight back. `eval` covers all of them.

`eval` alone, though, contains the error WITHOUT blanking the value, so
the name would be reported as blanked having never been blanked --
measured: attempted=59 with UID still 1000. So each name is verified
after the attempt and only counted when it actually blanked. A name that
could not be blanked currently falls into the "allowed" count, which is
imprecise in the safe direction; the honest third count ("tried and could
not blank") needs a report-format change -- readReport treats every
non-first non-blank line as a blanked name -- and belongs with #394.

Why the suite stayed green: EnvAllowListScrubTest starts zsh from a
CLEARED parent (`pb.environment().clear()`), where UID is not an
exported name at all, so the abort cannot reproduce there. The new test
supplies the production shape -- UID present and exported -- and asserts
a report exists, which is written by the last statement in the file and
so proves the script ran to completion. Verified by mutation in both
shapes: reverting the fix fails exactly that test and no other.

Orthogonal to #388, which adds a fifth call site to a script that still
aborts at the same name.
 2026-09-10 01:12:27 +00:00