Skip to content

OpenAI

Forced rm detection tightened in Codex

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →

OpenAI’s Codex team said on August 19 that changes rolled out over the previous couple of weeks have further reduced the risk of Codex performing destructive actions during its work. The team posted the summary in a recap on Twitter. The work started a few weeks earlier, when the team began investigating a small number of reports.

The recap doesn’t come out of nowhere. Earlier in the summer, users found GPT-5.6 Sol in Codex deleting real user files. A cleanup command meant for temporary working folders misused system variables such as $HOME, so it targeted the actual home directory instead of a temp path. The fix for that bug, as it was described at the time, had Codex verify deletion targets before running them, create fresh temporary folders, block full-access mode from being triggered by accident, and apply stricter checks to risky delete commands.

The change the team spells out in its GitHub post is stronger forced rm detection. It now picks up literal forced rm invocations inside control flow, substitutions, pipelines, traps and nested shells, as well as behind sudo or env wrappers. The detection is tied to the literal invocation, which leaves the wrapped and nested cases as the part that actually got wider.

OpenAI also says that most Codex Desktop token usage inside the company comes from Auto-review mode.

For users, OpenAI’s recommendation is to stay in a sandboxed mode and keep the app updated.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.