If you use OpenAI's Codex on an individual account, check two settings today: Improve the model for everyone and Codex's separate Include environments. A developer posted on Hacker News on November 3, 2026 that their team had turned data sharing off, and that "yesterday my coworkers and I noticed that the setting had been silently reenabled in Codex." OpenAI has not confirmed this, so this post separates what is claimed from what OpenAI's own documentation says.
This matters to people running coding agents because a Codex session is not a casual chat. It can contain source code, file paths, error output and, occasionally, pasted credentials. Whether that content is eligible for model training is a setting you control, if it stays where you put it.
TL;DR
| Question | Short answer |
|---|---|
| Am I affected? | Only if you use Codex on a Free, Plus or Pro account and the training setting is on. Check it rather than assume. |
| Is the reset confirmed? | No. It is a user report, with an earlier similar thread. We found no OpenAI statement. |
| Is there a fix? | Nothing to patch. Turn the setting off and re-check it periodically. |
| Does data leave my machine? | Yes, by design: Codex sends your prompts and context to OpenAI. The setting decides whether eligible content can train models. |
| Are Business, Enterprise and API accounts included? | OpenAI says these are excluded from training by default. |
What is claimed
The post, titled "OpenAI seems to be silently re-enabling model training on Codex chats," makes three claims. The author says their company is careful about leaking intellectual property to frontier labs and that everyone had disabled data sharing in Codex and other agents. They say the setting had been re-enabled for several people on the same day. And they say none of them knowingly opted back in.
The author points readers to the ChatGPT data controls page at chatgpt.com/settings/data-controls and the toggle named "Improve the model for everyone." They also note that the same thing was raised in an earlier Hacker News thread from September 10, 2026, titled "Tell HN: OpenAI keeps re-enabling the 'allow training' setting."
That is a small sample from an anonymous poster. It is not telemetry, and it does not show a mechanism. No screenshots, timestamps or account types were attached in the text we read.
What OpenAI documents
OpenAI's help article, How your data is used to improve model performance, was marked as updated 14 days before we read it. It says the following, which we paraphrase closely:
- For individual services (Free, Plus, Pro), OpenAI may use your content to train its models. To opt out, turn off Improve the model for everyone under Settings > Data controls in ChatGPT, or select Do not train on my content in the Privacy Portal.
- "Either option is sufficient for ChatGPT conversations and Codex tasks. You don't need to do both."
- After you opt out, OpenAI says it won't use your new conversations to improve its models.
- Codex "has a separate Include environments setting in Codex settings for allowing training on full environments. Changing your settings in ChatGPT or the Privacy Portal does not change that setting."
- If you give feedback with thumbs up or thumbs down, the entire conversation may be used to train models even if you opted out.
- By default, ChatGPT Business, Enterprise, Edu and the API are not used to improve models. See OpenAI's enterprise privacy page.
Two details are easy to miss. First, there are at least two Codex-relevant controls, and flipping the ChatGPT one does not touch Include environments. If someone audited only the ChatGPT toggle, they may have missed the other. Second, the opt-out is forward-looking in OpenAI's wording.
Claimed vs verified
| Claim | Status | Source |
|---|---|---|
| The "Improve the model for everyone" toggle was found on after being turned off | Reported by one poster, unconfirmed | Hacker News, Nov 3, 2026 |
| It happened to several coworkers at once | Reported by the same poster, unconfirmed | Same thread |
| The same thing has happened before | Reported in a September 2026 thread, unconfirmed | Hacker News, Sep 10, 2026 |
| Opting out covers Codex tasks | Documented by OpenAI | OpenAI help center |
| Codex has a separate environments setting that the toggle does not change | Documented by OpenAI | OpenAI help center |
| OpenAI deliberately re-enables the setting | No evidence | Speculation in comments |
| Business, Enterprise, Edu and API are excluded by default | Documented by OpenAI | OpenAI help center |
We did not reproduce the reset. We could not log in to the affected accounts, and we have no evidence of its cause. A bug, a migration of settings, a new consent flow, or a user-interface problem are all plausible explanations, and none has been shown.
What developers are saying
The discussion has substance in places, though much of it is opinion.
- The poster edot says that turning on Advanced Account Security, which requires hardware keys, also locks data sharing off. A commenter in the earlier thread, arpinum, makes the same claim. We could not verify this in OpenAI's documentation, so treat it as a user report.
- vibranium suggests signing a zero-data-retention contract or using Bedrock instead of relying on a toggle. CEOatMindfries suggests that coding agents should have a layer that keeps IP from leaking, handling it deterministically like environment variables.
- In the September thread, antonok reports that the toggle writes a local storage entry whose value did not seem to matter on new tab loads, and wondered whether this was a user-interface bug. That is one person's inspection, not a finding.
- cbg0 reports that their opt-out held for months, and andsoitis says they have not seen the reset. So the experience is not universal.
- amelius suggests documenting your own settings, since a record of what you configured helps if there is a dispute.
What this means if you run AI coding agents
These steps come from OpenAI's help article and common sense, not from a confirmed bug.
- Check the ChatGPT toggle. Open Settings > Data controls and look at "Improve the model for everyone." Turn it off if you do not want eligible content used for training.
- Check Codex separately. Open the Codex settings and look at Include environments. OpenAI says the other controls don't change it.
- Consider the Privacy Portal option. "Do not train on my content" is documented as sufficient for Codex tasks, though the separate environments setting remains.
- Record what you see. Take a dated screenshot or note per person on the team, and re-check weekly or after any app update. If it flips, you have evidence to send to OpenAI support.
- Avoid the feedback buttons on sensitive sessions. Per OpenAI, thumbs up or down can make the whole conversation eligible for training.
- Use the right account type for proprietary code. A managed workspace or API usage has no-training defaults according to OpenAI. Check the contract terms yourself.
- Keep secrets out of reach. A setting protects against training use, not against exposure in transit or in logs. Credentials in
.envfiles, shell history or pasted output should not be readable by the agent in the first place. See our guide on how API keys leak through logs and environment variables. - Rotate anything that may have been sent. If a real key appeared in a session while training was on, treat it as exposed and rotate it.
What Beam can and cannot do here
Beam is local, on-device monitoring for AI coding agents; its local-first design is described in the introduction. It can help with the part of this story you control: seeing what the agent reads and runs, flagging credential exposure in tool calls, and redacting known credential formats before anything is stored on your machine. Nothing leaves your device by default.
It cannot see or change your OpenAI account settings, cannot tell whether the training toggle is on, and cannot stop Codex from sending a prompt to OpenAI. Its deny is hook-based for Claude Code and Codex PreToolUse, not a network filter or kernel sandbox, and its scanning is heuristic. Think of it as a way to reduce how much sensitive material the agent touches in the first place, which is the only protection that holds regardless of any toggle. For another case of agent traffic going somewhere unexpected, see our write-up on unauthorized Codex and Claude Code relays, and for a related look at how labs monitor Codex sessions, see the METR and Redwood agent-monitoring report.
What we will watch
We will update this post if OpenAI confirms or denies a reset, publishes a cause, or changes the controls. Things worth watching for are a statement from OpenAI support, a change to the help article, a new consent prompt in the Codex app, and independent reproductions with dates and account types. If you have a reproducible case, a dated screenshot is more useful than a thread comment.
The broader lesson is not specific to one vendor. Privacy settings on a hosted agent are a policy you have to audit, not a state you set once. The same team that turns data sharing off should write down who checks it, how often, and what they do when it changes. For related reading on privacy around chat products, see OpenAI's Project Lily and human review of chats.
Frequently asked questions
Did OpenAI silently re-enable training on Codex chats?
That is a claim, not a confirmed fact. A Hacker News poster says that they and their coworkers found the "Improve the model for everyone" setting switched back on after they had turned it off, and an earlier thread from September 2026 reports similar experiences. OpenAI has not, as far as we could find, confirmed a reset or explained a cause.
Which settings control training for Codex?
OpenAI's help page says that for Free, Plus and Pro accounts you can opt out by turning off "Improve the model for everyone" in ChatGPT Data controls or by choosing "Do not train on my content" in the Privacy Portal, and that either is sufficient for ChatGPT conversations and Codex tasks. Codex also has a separate "Include environments" setting for training on full environments, which those two controls do not change.
Does opting out remove data that was already used?
OpenAI's page says that after you opt out, it will not use your new conversations to improve its models. It does not say that content already used is removed, so treat anything sent while the setting was on as potentially included.
Do Business, Enterprise and API accounts have this problem?
OpenAI states that by default it does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or the API to improve its models. The reports we found concern individual accounts, so a managed workspace follows different rules.
Can giving feedback undo my opt-out?
Yes for individual plans. OpenAI says that if you give a thumbs up or thumbs down, the entire associated conversation may be used to train its models even if you opted out. For a coding agent that conversation can include code, file paths and tool output.
What should I do if my team cannot risk sending source code?
Do not rely on a single toggle. Check the settings regularly, use a managed workspace or API terms with no-training defaults where possible, keep secrets out of the agent's reach, and watch what the agent reads and sends from your own machine.
Related reading
- Introducing Beam: local-first AI agent security
- How API keys leak through logs and environment variables
- OpenAI's Project Lily and human review of ChatGPT chats
- METR and Redwood on monitoring OpenAI agents
- Beam on the web: what it watches
Details reflect the OpenAI help article and the Hacker News threads as we read them on November 3, 2026. OpenAI's settings and wording can change; Beam is a local prototype, so check current documentation before acting.