OpenAI Fixes Codex File Deletion Issue with New Safeguards

OpenAI's Codex team lead Thibault Sottiaux responded on X about the issue where Codex, using GPT-5.6 models, accidentally deleted some user files. The problem happened when a cleanup command wrongly targeted the user's home directory, and the model didn't check what was there before deleting. Sottiaux said they've added multiple safeguards: Codex now checks before deleting, creates new temp folders, avoids reusing system variables, and stops if unsure. OpenAI also improved execution checks to catch risky delete commands, made it harder to enable full access without clear warnings, updated auto-review tools, and built models to reproduce the failures. They also filtered out destructive behavior from training data and added extra training tasks.