# OpenAI Silently Re-Enabled Codex Training: Fix It Now > Source: > Published: 2026-11-10 11:08:33+00:00 The model training toggle you switched off in Codex may be on again. On November 3, a Hacker News thread confirmed what several developers had suspected: OpenAI silently re-enabled the “Improve the model for everyone” setting for Codex users who had explicitly opted out. As of November 10, OpenAI has not acknowledged the incident or explained why it happened. ## There Are Two Settings. Most Developers Only Know One. This is where the story gets worse. Codex has two separate training controls, and they operate independently of each other. Disabling one does not disable the other. - **Toggle 1 — ChatGPT Data Controls:** Navigate to`chatgpt.com/settings → Data Controls → “Improve the model for everyone”` and switch it off. This is the setting most people know about. - **Toggle 2 — Codex “Include environments”:** A separate control inside Codex settings that governs whether your full development environment — not just your chats — is used for training. Changing your ChatGPT settings does not touch this one. As [DeveloperDigest noted](https://www.developersdigest.tech/blog/codex-training-setting-reenabled-include-environments-check-2026): “you can have done the ‘right’ thing in ChatGPT and still have a second control governing your repo environments.” If you have only ever disabled Toggle 1, your code environments may still be feeding the model. ## What Happened on November 3 A Hacker News post drew 450+ upvotes and 103 comments from developers reporting the same experience: their training opt-out had been silently reverted. One developer noted their account history showed two previous requests to disable training — both had been switched back on. Another managing monitoring infrastructure framed it plainly: “a pretty serious issue since we tell our customers that none of their data is being trained on.” Nobody has identified a reproducible trigger. It is unclear whether the re-enablement came from a platform update, a migration, or a bug. [OpenAI has not issued any statement](https://emergent.sh/news/openai-codex-training-toggle-silently-re), changelog entry, or public acknowledgment. ## Who Is and Is Not Affected The affected tiers are **Free, Plus, and Pro personal plans**. On these, training is on by default and requires manual opt-out — these are also the plans most individual developers use for day-to-day work. **Not affected by default:** ChatGPT Business, Enterprise, Edu, and the OpenAI API. On those tiers, training is excluded without user action. The shadow usage problem amplifies the risk. Developers routinely use personal accounts — running under consumer terms — for work involving proprietary code. When the training toggle gets re-enabled, company IP can flow into training pipelines without anyone in the organization noticing. [Analysis of enterprise shadow AI usage](https://www.promptzone.com/rohan_murphy/openais-silent-codex-policy-shift-what-developers-need-to-know-1nkg) puts OpenAI at 53% of tracked activity — making it the single largest point of exposure. ## Toggles Are Not Enough Even fixing both settings today leaves gaps. Changes are not retroactive — code submitted under an enabled training setting may already be flagged. Giving feedback in Codex (thumbs up or down) overrides opt-out on personal plans. And since the trigger for re-enablement is unknown, there is no guarantee your settings will hold after the next platform update. AI Weekly’s recommendation is blunt: “verify toggle state across every Codex seat today and treat the UI flag alone as insufficient evidence for contractual claims.” For teams with compliance requirements, that is a reasonable posture. ## What to Do Right Now Fix both settings, then address the structural problem: 1. Go to **chatgpt.com/settings → Data Controls → “Improve the model for everyone” → OFF** 2. Open **Codex settings → “Include environments” → OFF** 3. If you use Codex for work code on a personal plan, move to a **Business or Enterprise workspace** — training exclusion is the default there, not a toggle you have to remember to maintain 4. If you access Codex via the **OpenAI API** , you are protected by default; API data is not used for training unless you explicitly opt in For context: Anthropic’s API uses opt-in by default for training, and [GitHub Copilot offers enterprise on-premise options](https://docs.github.com/en/copilot/managing-copilot/managing-github-copilot-in-your-organization/managing-policies-for-copilot-in-your-organization) for teams with the highest compliance requirements. The asymmetry between OpenAI’s consumer and enterprise plans is by design — but that design only holds if developers know which tier they are on and what it means for their data. OpenAI’s silence on this incident is worth noting. When a privacy control fails quietly and the company says nothing, the burden falls entirely on individual developers to audit their own settings state. That is not a reasonable posture to hold your users to — especially when those users are making contractual commitments to their own customers based on what that toggle says.