Turned Off OpenAI Training? Codex Has a Second Switch You Probably Missed.

OpenAI's own help page says the main opt-out doesn't cover context from Codex environments, and sessions in its Daybreak security program can be trained on anyway.

The obvious way to keep work out of OpenAI's training is one setting: "Improve the model for everyone." OpenAI's own help page says that is not the whole job for Codex, its coding agent. In a section on Codex tasks, the data-controls article (as archived November 13) says Codex "has a separate Include environments setting" that controls whether extra context from Codex environments can be used for training. "Changing your ChatGPT setting or opting out through the Privacy Portal does not change that setting." A Codex environment, per OpenAI's Codex docs, is the cloud container a Codex task runs in: your repository checked out, plus the setup script, tools and settings you give it.

The hosts are AI characters.

Listen to this story
1:54

Jon's agents run Codex workers all day, so this is a setting that matters to his own work, not a policy footnote.

The same page carries a second exception. On a personal account, sessions in Daybreak, OpenAI's program for sensitive security work, may be used to improve models "even when 'Improve the model for everyone' is off." The Hacker News reports that this starts November 29, 2026. Users pushed back at once: a security researcher called it "cartoonishly evil", and a thread on OpenAI's own forum asks whether it covers work sent before that date. Meanwhile OpenAI's Advanced Account Security page, a setting the Daybreak program requires, says conversations "are not used to train OpenAI models" while it is on. OpenAI has not said how the two pages fit together.

There is also a louder claim going around: that OpenAI quietly turns the main setting back on. Two Hacker News threads, one in September and one in November, carry users reporting just that. Several people in the same thread checked and found theirs still off, and nobody has shown a cause. The documented gap is the part to act on.

Close both doors

  1. Main setting. In ChatGPT, open Settings, then Data controls, and turn off "Improve the model for everyone." On Free, Plus and Pro personal workspaces it starts on.
  2. Codex setting. Open Codex settings (on the web, chatgpt.com/codex/settings, the same place OpenAI's docs send you to set up environments) and turn off "Include environments." The ChatGPT switch and the Privacy Portal do not touch it.
  3. Daybreak. On a personal account, keep client code out of Daybreak sessions, or wait for OpenAI to say what happens to work sent before November 29.
  4. Check again later. Given the reports, look at both settings once a month.
  5. Mind which workspace. The help page's "on by default" line covers personal workspaces on Free, Plus and Pro. For client code, check which workspace a Codex task runs in before it starts.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.