Daybreak Blue training: OpenAI's opt-out exception for security work
The Morse teamPublished 8 min read
As of November 2026, OpenAI may train on Daybreak Blue sessions on a personal ChatGPT account even with "Improve the model for everyone" switched off. Its Data Controls page says that "during Daybreak sessions on your personal account, we may use content from your interactions to improve our models even when 'Improve the model for everyone' is off." The page gives no start date. A Daybreak Blue user reported an OpenAI email putting the training change at November 29, 2026, and OpenAI hasn't published a separate opt out.
We read OpenAI's Data Controls page, its Daybreak overview, troubleshooting, and Codex CLI pages, its Advanced Account Security page and launch post, its model-improvement page, and its Health page on November 15, 2026, along with the Internet Archive's copies of the Data Controls page. OpenAI deserves credit for putting the exception on its main privacy page in plain words, and for saying your setting "still applies to your other activity when not in a Trusted Access session." What's only reported, we label as reported.
What is ChatGPT Daybreak?
Daybreak is OpenAI's Trusted Access for Cyber program: approved security people get fewer refusals from OpenAI's models for "authorized defensive cybersecurity work," according to OpenAI's Daybreak overview. It has two tiers.
- Daybreak Blue gives "reduced refusals" on mainline models such as GPT-6 Sol and GPT-6 Luna, for work like vulnerability triage, malware analysis, and incident response. Individuals can apply for it.
- Daybreak Red adds cyber models such as GPT-5.6 Cyber, for penetration testing and exploit development. OpenAI says it's "available only to approved business and enterprise organizations," though "some individuals may have access as an exception."
To get Blue as an individual, you need "an eligible paid plan, Advanced Account Security, and at least one compatible FIDO2 hardware security key," plus identity verification. Daybreak doesn't "grant Zero Data Retention by default." Our guide to AI for security research compares Daybreak with Anthropic's and Google's programs.
Does OpenAI train on Daybreak Blue sessions?
It may, on personal accounts, whatever your training switch says. OpenAI gives its reason in one line: "This helps us to strengthen safeguards against cyber misuse."
The wording is broader than Blue. The paragraph is titled "Trusted Access on personal accounts," names Daybreak Blue only as an example ("Trusted Access programs like Daybreak Blue"), and then speaks of "Daybreak sessions" in general. It doesn't say whether files, tool output, or older content pulled into a session count, how long the content is kept, or whether people read it.
In Codex, a "session" is something you switch on. OpenAI's Codex CLI page says turning it on with /daybreak "will apply to the active thread and be saved as the default for new sessions," that daybreak = true in config.toml makes it the default, and that "resumed and forked sessions retain their saved Daybreak choice." In the Codex app, the Daybreak toggle "is off by default."
When does the Daybreak training change start?
OpenAI's help pages give no date, and the November 29 start comes from an email. On October 30, 2026, a Daybreak Blue user wrote on OpenAI's community forum that an email said "from November 29 onward, OpenAI may use content from Daybreak Blue sessions on personal accounts to improve their models." We haven't seen the email, so treat November 29 as reported. Another user on the thread asked whether the change reaches content sent before that date, and OpenAI's pages don't say.
The help page changed the same day. The Internet Archive's copy from 15:35 UTC on October 30 has no Daybreak paragraph, and its copy from 19:18 UTC has it, marked "Updated: 1 hour ago." The page as we read it on November 15 carries the same wording, with no date in it.
Is there a separate Daybreak opt out?
Not on OpenAI's pages as of November 15, 2026. RuntimeWire reported on October 30 or 31 that a second notice told personal-account Blue users they would get "a separate way to opt out of model improvement for Daybreak sessions," that interactions flagged for safety review may still be used for safety purposes, and that more detail would come closer to November 29. We couldn't find that notice on any OpenAI page.
The Data Controls page still describes only the exception, and OpenAI's model-improvement page lists one way to reach training after you opt out, the thumbs buttons, and doesn't mention Daybreak at all.
Doesn't Advanced Account Security turn training off?
OpenAI's pages disagree on this. Daybreak requires Advanced Account Security, and OpenAI's Advanced Account Security page says, with no exception listed: "Conversations are not used to train OpenAI models while Advanced Account Security is enabled." OpenAI's launch post called this "Automatic training exclusion" and said "that preference is automatic."
So individual Daybreak users had to turn on a setting that OpenAI said kept their conversations out of training. Then the Data Controls page carved Daybreak sessions out of the opt out. As of November 15, neither the Advanced Account Security page nor the model-improvement page carries the Daybreak exception. Coverage we read, such as explainx, sourced the security-setting promise to a Hacker News comment and said it couldn't confirm it in OpenAI's documentation, though both OpenAI pages linked above carry it. We first noted the clash in does Codex train on your code.
Which accounts does the Daybreak change leave out?
The exception names personal accounts only. OpenAI's Data Controls page says it doesn't use content from "ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces" to train by default, nor "inputs and outputs provided through the API Platform."
Two other rules stand on a personal account. Temporary chats "are not used to improve OpenAI models" while they stay temporary, and rating a reply means "the entire conversation associated with that feedback may be used to train OpenAI models," opt out or not (OpenAI). Everything else about the switch is in does ChatGPT train on your data.
The Daybreak paragraph doesn't mention Health. OpenAI's Health page, read November 15, still says connected records and the chats that use them aren't used to train its foundation models "regardless of the model-training setting you choose," and our page on whether ChatGPT Health is private covers the rest.
Which training rule applies to you?
Each row comes from OpenAI's own pages, read November 15, 2026.
| Your situation | Can OpenAI train on it? | Source |
|---|---|---|
| Personal plan, "Improve the model for everyone" on | Yes, by default on Free, Plus, and Pro | OpenAI |
| Personal plan, switch off | No, for new conversations | OpenAI |
| Switch off, but you rate a reply | Yes, the whole conversation | OpenAI |
| Advanced Account Security on, no Daybreak | No, "while Advanced Account Security is enabled" | OpenAI |
| Daybreak session on a personal account | Yes, "even when 'Improve the model for everyone' is off" | OpenAI |
| Business, Enterprise, Edu, or Healthcare workspace | No, by default | OpenAI |
How did the Daybreak training change unfold?
| Date | What happened | Source |
|---|---|---|
| April 30, 2026 (earliest archived copy) | OpenAI introduces Advanced Account Security: "conversations from those accounts will not be used to train our models" | OpenAI |
| June 1, 2026 | Individual Trusted Access members using "our most cyber capable and permissive models" must enable Advanced Account Security | OpenAI |
| October 1, 2026 | Existing individual Daybreak users must have an eligible paid plan, Advanced Account Security, and a FIDO2 hardware key "to keep access" | OpenAI |
| October 30, 2026 | The Data Controls page gains "Trusted Access on personal accounts" between 15:35 and 19:18 UTC | Internet Archive |
| October 30, 2026 | A Blue user reports an OpenAI email giving November 29 as the start (reported) | OpenAI forum |
| October 30 to 31, 2026 | A second notice reportedly promises a separate Daybreak opt out (reported) | RuntimeWire |
| November 15, 2026 | OpenAI's help pages show the exception, no separate opt out, and an Advanced Account Security page that still promises no training | OpenAI |
| November 29, 2026 | Reported start of training on personal Daybreak sessions | OpenAI forum |
How do you keep security work out of OpenAI's training?
- Turn Daybreak off for anything you can't share. Your setting "still applies" outside a Trusted Access session. In Codex CLI, run
/daybreakagain to switch it off, use-c daybreak=falsefor one run, and checkconfig.toml, since a saveddaybreak = truecarries into new sessions. - Leave client code and unreported bugs out of Daybreak sessions until OpenAI publishes the opt out it reportedly promised, and read your engagement terms before pasting anything a client owns.
- Skip the thumbs buttons. A rating opens the whole conversation to training, Daybreak or not.
- Keep a dated copy of the Data Controls page and of any email OpenAI sends you, since the help page gives no date.
- Dropping Advanced Account Security isn't a way out. OpenAI's troubleshooting page says turning it off can cost you Daybreak approval, and once it's off, only your regular switch keeps chats out of training. Losing Trusted Access "does not delete your ChatGPT account or existing chats."
This isn't legal advice.
How Morse keeps your security work out of training
On ChatGPT, your opt out is a setting on OpenAI's servers, and OpenAI can add an exception to it with one paragraph on a help page. Morse doesn't train AI on your chats, so there's no switch for a later program to override. Your chat history is stored in your browser, encrypted, not on Morse's servers, and Morse's servers don't keep or log what you write.
- E2EE, for anything no one should read. Your browser checks the enclave's hardware attestation, then encrypts the message so only that enclave can open it. Morse's servers relay ciphertext they can't read, and each reply has a lock that shows every check that passed. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models like GLM, Kimi, and DeepSeek, with no web search or image generation.
- Private, for routine security questions you'd rather no host kept. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- Anonymous, to ask GPT without the provider knowing it's you. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider still sees the words and may keep them for a limited time to check for abuse, under its own policy, so leave client names and hostnames out.
On paid plans, picking an unfiltered model gives answers other apps refuse, including help testing your own security. Morse doesn't sell your data or show ads, and because it doesn't store chats, it usually has very little to give authorities. Clearing your browser or losing the device deletes the chats and Morse can't recover them, so keep your own copy of findings you'll need.
Start free, or send a first question from the Morse home page with no account. E2EE chats come with Pro, at $19 a month.
Common questions
Is there a separate opt out for Daybreak training?
Not on OpenAI's help pages as of November 15, 2026. RuntimeWire reported that a second OpenAI notice promised one, but the Data Controls page still describes only the exception, and your normal setting covers only activity outside a Daybreak session.
Does the Daybreak change apply to ChatGPT Business or Enterprise?
The exception names personal accounts. OpenAI says it doesn't use content from Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces, or from the API Platform, to train its models by default.
Does it cover Daybreak chats from before November 29?
OpenAI's pages don't say. A user on OpenAI's forum asked exactly this, and the help page gives no date at all, so the November 29 start is only reported.
If I leave Daybreak, do I lose my chats?
No. OpenAI's troubleshooting page says losing Trusted Access does not delete your ChatGPT account or existing chats. Turning off Advanced Account Security can cost you Daybreak approval.