Developers say OpenAI's Codex turned model training back on for accounts that had already opted out, and the claim spread across Hacker News threads on November 3, 2026. The reports describe a Codex training setting that flipped on without notice, a problem for anyone who sends proprietary code or client work through the tool. OpenAI has not publicly confirmed or explained the change in any source reviewed for this story, so the details below are reported, not established.
What the reports say happened
The most-discussed thread on Hacker News carries a title suggesting OpenAI is silently re-enabling model training on Codex chats. Users in that thread describe training toggles that were back on despite earlier opt-outs. One commenter says two past opt-out requests had been granted, and another says the setting was flipped back on after they had turned it off and other team members noticed the same change. That commenter also said they built a tool to flag agent accounts with training enabled.
A separate Ask HN post from the same day described a team that had disabled data sharing across Codex and Claude Code, only to find the setting switched on for several coworkers. That is a single anecdote, not a dataset.
A summary from Emergent says Codex silently re-enabled data training on November 3, 2026, despite users' prior opt-outs, and that OpenAI had not acknowledged the change. Emergent is a secondary account, so its claims trace back to the same forum evidence rather than to company documents.
Engagement figures vary by report. One summary cites a 345-point thread, while a separate discussion is cited at 193 points. These are different threads measured differently, not one total.
Which Codex settings and surfaces may be affected
Sources do not agree on where the change lands, and that is the first thing to pin down before changing any account. A PromptZone guide says the change applies to chats in the ChatGPT interface that use the Codex model, and that the opt-out appears to be opt-out rather than opt-in. Developers Digest points to two switches instead: the main training toggle and a separate Codex setting that covers training on full environments, which it draws from a guide written in 2025.
The table below summarizes the settings readers are being told to check. Confirm the current labels before you act, because the sources describe them from different points in time.
| Setting | Where sources place it | What to verify |
|---|---|---|
| Main training toggle | ChatGPT data controls, labeled "Improve the model for everyone" per PromptZone | Current label and whether it is on |
| Codex environment training | Separate Codex switch, per a 2025-era guide cited by Developers Digest | Whether it still exists and its state |
| Privacy portal "do not train" | Described as equivalent for the main toggle | Whether it also overrides the Codex switch |
The problem may also be older than November. A September Hacker News post reported resetting the same toggle more than once and finding it re-enabled. That post dates from about two months before this coverage, which suggests a pattern worth investigating, though one post does not prove one.
What the evidence cannot yet show is scale. Every report reviewed here is anecdotal forum material. Nothing in the research establishes how many accounts were affected, whether the change was deliberate, or whether it reached only a handful of workspaces.
Why a reverting opt-out is a trust problem
Opt-out controls exist because many people assume that a setting they have turned off stays off. That assumption is the core of this story. Nothing reviewed here shows a data breach, and the reports concern settings, not a leak of stored code. But a control that appears to revert undermines the basis on which people share sensitive material.
The stakes differ by user type. An individual developer or freelancer on a personal ChatGPT or Codex account may have code or client data used for training if the setting was on. A company with a security team faces a different question: whether training was enabled on any account that touched company code, and what that means for compliance and audit obligations. The Emergent summary points to exactly those compliance and audit questions for companies using Codex on company code.
OpenAI's enterprise and API tiers are described as excluded from training by default. Developers Digest says OpenAI states that API, Business and Enterprise data is not used for training by default. That is why several sources steer work users toward managed workspaces rather than personal accounts. The research does not establish whether team-level settings override individual opt-outs, so admins should treat that as an open question.
The episode also lands amid wider concern about AI tools acting on users' behalf. Apple's SEC filing warning that AI agents could act without permission raised a related concern from a different angle: a setting users assume governs behavior has to hold up in practice.
How to check your Codex training setting now
Work through these steps on every account that touches code or client data, not just the one you use most.
- Open ChatGPT data controls and find the training toggle. PromptZone describes the main control as labeled "Improve the model for everyone," but confirm the current label before relying on it.
- Check for a separate Codex setting that covers training on full environments. Developers Digest says this switch sits apart from the main toggle, so turning off one does not prove the other is off.
- Test the privacy portal "do not train" option. Developers Digest says it is equivalent for the main toggle, but one commenter reports it may not override a separate Codex setting. Verify both after you change them, and check again in a few days.
- Keep secrets out of prompts and environment files. API keys, passwords and credentials should never reach a model, whatever the settings say.
- For work, use a managed Business or Enterprise workspace or the API. Sources present this as the structural fix, since those tiers are described as excluded from training by default.
- Write down what you find, with dates, for each account. If you run a team, that record is the starting point for any audit.
What OpenAI has said, and what remains unknown
As of November 6, the searches behind this story found no official OpenAI statement about the reports. The Emergent summary says OpenAI had not acknowledged the change, and no on-record quote from OpenAI or a named official was found. Readers should not assume the company has explained the behavior, or that it has denied it.
Several basic questions remain open. Developers Digest lists a bug, a migration or user error as possible causes, but the research does not establish which one applies, or whether the change was a deliberate policy shift. It is also unclear whether data already used for training can be removed, whether the opt-out respects enterprise or team settings, and whether any regulator or state attorney general is examining the matter. None of those answers appeared in the research.
What to watch next
The most useful next step would be an official OpenAI statement that names the affected surfaces, says whether the Codex environment setting is separate from the main toggle, and explains what happens to data collected since November 3. Without that, developers will keep relying on forum threads and third-party guides to understand their own accounts.
Watch whether the September reports and the November reports share a root cause. A repeating pattern would move this from a single glitch toward a question about how OpenAI tests and records user preferences. Watch too whether the privacy portal opt-out holds across both switches, since that is the first fix most users will try.
Readers tracking OpenAI's broader security questions may also want USA 24's coverage of OpenAI's autonomous AI pipeline, which built 948 Linux root exploits. The two stories are separate, but both test how much users can trust what the company says its systems do.
