OpenAI has shipped a security update for Codex after multiple users reported that GPT-5.6 Sol was autonomously deleting files while running. Not temp files. Actual user data.
The root cause, according to OpenAI, was a cleanup command gone rogue. Codex was supposed to clear out temporary working files, but it started using system variables like $HOME for those temp folders. A faulty delete command then pointed at the real home directory instead — and wiped it. That’s the kind of bug that makes developers start sweating. You don’t expect an agent to nuke your home directory because it mixed up its workspace paths.
The fix layers in several safeguards. Codex now has to verify deletion targets before executing them, creates fresh temporary folders instead of borrowing system paths, and stops misusing system variables altogether. OpenAI also added stricter checks to catch risky delete commands, and full-access mode can no longer be triggered accidentally. The company recommends users stick with one of the sandbox modes and keep the app updated.
This is a reminder that autonomous coding agents are still, at their core, executing shell commands on your machine. The difference between a useful assistant and a destructive one is often a single path variable or a missing confirmation step. OpenAI’s response was relatively quick, but the fact that this shipped in the first place suggests agentic tools need much more rigorous guardrails around filesystem operations — especially as more developers hand over write access without thinking twice.
