Allow DYLD_* in codex's shell

Resolved 💬 2 comments Opened Dec 23, 2025 by amonks Closed Feb 23, 2026

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex is running?

codex-cli 0.77.0

What subscription do you have?

usage tier 5

Which model were you using?

gpt-5.1-codex

What platform is your computer?

Darwin 25.1.0 arm64 arm

What issue are you seeing?

I'm working on something where I need to set DYLD_* environment variables. It looks like the process hardening crate prevents this.

What steps can reproduce the bug?

prompt ! DYLD_TEST=yes env | grep DYLD_TEST
with shell_environment_policy.set = { DYLD_TEST = "yes" }, prompt ! echo $DYLD_TEST

What is the expected behavior?

in both cases, I expect to see DYLD_TEST set to "yes", but it is unset.

Additional information

I understand not allowing DYLD_* (or LD_PRELOAD or whatever) in codex's own processes, but not allowing them in the shell seems unreasonable. Lots of tasks depend on these tools.

View original on GitHub ↗

2 Comments

johnzfitch · 7 months ago

Thank you @amonks this issue has a bigger surface area than this and I'm glad some others are starting to file reports too.

Original PR that introduced the problem:

PR: #4521
Commit: b8e1fe60c5af4439c7e159a49ba11a0ef3f7eb7c
Date: September 30, 2025
Author: Michael Bolin <mbolin@openai.com>
Title: "fix: enable process hardening in Codex CLI for release builds"
Rationale: "I don't believe there is any upside in making process hardening opt-in"

Why the original rationale was incorrect:

CUDA development is not "tinkering" - it's mainstream ML/AI work
"Build from source" is not viable for most enterprise/academic users
Performance regression (11-300x) is severe and unacceptable
Security benefit is minimal beyond existing Landlock/Seatbelt boundaries

This is not a new feature request. It is a request to restore the previous opt-in behavior from commit 0dd49683, which was reverted by commit b8e1fe60.

etraut-openai contributor · 6 months ago

This feature request hasn't received enough upvotes, so I'm closing it.