OpenAI told members of Daybreak Blue, its vetted program for cybersecurity work, that from November 29 it may use their sessions on personal ChatGPT accounts to improve its models, even when the "Improve the model for everyone" setting is off. Hours later, a follow-up email shared by several recipients promised a separate opt-out, with one exception: interactions flagged for safety review may still be used.
The change matters because of what people bring to those sessions. OpenAI's own Daybreak overview lists vulnerability triage, malware analysis, incident response and secure code review as the work Daybreak Blue is for, on systems the user owns or is authorized to test.
What OpenAI said, and when
The first email, which recipients began posting on X at about 7 p.m. UTC on October 30, notes that Daybreak requires Advanced Account Security, "which turns off the 'improve the model for everyone' setting." It continues: "However, from November 29, we may use content from Daybreak Blue sessions on personal accounts to improve our models, including to help strengthen safeguards against cyber misuse."
The same policy is now on OpenAI's help center. Its data controls page, updated on or about October 30, has a new section on Trusted Access:
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.OpenAI, Data controls in ChatGPT
That page gives no start date, and when SaasBuzz read it at 6:20 p.m. UTC on October 31 it did not mention an opt-out. The opt-out appears so far only in the second email. Costin Raiu, a threat researcher on the team of the threat intelligence firm TLPBLACK, posted its text early on October 31 and asked how OpenAI would use the flagged sessions:
Post on X by @craiu
Some Daybreak users are receiving emails from OpenAI that their can opt out of training, except for cases that trigger the safeguards. Hopefully this will be used to reduce the number of false positives with the guardrails, not the other way around? OpenAI: “We apologize for the confusion of the previous communication. Even past November 29th, Daybreak Blue users will have a separate option to opt out of Daybreak-enabled sessions to improve models. Given the advanced capabilities made available through this program, interactions flagged for safety review may still be used for safety purposes like improving our protections against these capabilities being used for harm. We will provide more information about how this system works closer to November 29.”
OpenAI had not posted the follow-up on its help center or its own X accounts by the time of writing, so the opt-out rests on the emails recipients shared.
Who is affected
The first email limits the change to Daybreak Blue sessions on personal accounts. It says the change "does not impact business or enterprise customers, or non-Daybreak Advanced Account Security sessions or other use of ChatGPT," and does not change OpenAI's handling of Zero Data Retention.
Account | Daybreak sessions in training |
|---|---|
Personal, Daybreak Blue | From Nov. 29, unless the promised opt-out is used; flagged chats excepted |
Business or Enterprise | Not affected, per the email |
Personal, outside Daybreak | The training setting still applies |
Sources: OpenAI's email of October 30 as posted by recipients; OpenAI's data controls page.
That split matches OpenAI's general defaults. Its help article on model training says that by default it does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or its API to improve its models, and its data controls page says that on personal Free, Plus and Pro accounts data sharing is on by default.
A second change reached at least some users. An X user who posts as Adolan (@adolandev) wrote that the copy of the email Adolan received also said individual users lose API access to Daybreak Blue on November 20, while Codex access through a ChatGPT sign-in stays. The post's screenshot shows the email's opening, which announces "two upcoming changes"; copies other recipients posted announce only the training change.
Date | What happens |
|---|---|
Oct. 30 | First email: training on personal Daybreak Blue sessions |
Oct. 31 | Follow-up: separate opt-out promised |
Nov. 20 | Individual API access ends, per one recipient |
Nov. 29 | Training may begin; details due before |
Source: OpenAI's emails as posted by recipients on X; SaasBuzz count from October 31: 20 days to November 20, 29 days to November 29.
Why researchers objected
The people the first email reached are the users who had done the most to lock their accounts down. Individual Daybreak access requires Advanced Account Security and a physical FIDO2 security key, which existing individual users had to have by October 1, and the email itself says that enrollment turns the training setting off. The notice then said the setting would not cover the sessions most likely to hold unpublished vulnerabilities and client code.
Alex Ellis, founder of the open-source project OpenFaaS, quoted the first email and said he could not see how he would keep using the model:
Post on X by @alexellisuk
What a shame from @OpenAI - I don’t see how I can keep using the model with this change from November 29, we may use content from Daybreak Blue sessions on personal accounts to improve our models, including to help strengthen safeguards against cyber misuse. We will use Daybreak Blue interactions for these purposes, even when the “improve the model for everyone” setting is off.
The follow-up answers most of that, but not all of it. OpenAI has not said how a session gets flagged for safety review, who reviews it, or how a flagged session would be used beyond "safety purposes." Its own description of Daybreak Blue is reduced refusals for defensive work, so the sessions most likely to trip a safeguard are the ones the program exists for.
What security teams can do now
Find out who on the team uses Daybreak Blue through a personal ChatGPT account rather than a company workspace.
Move client or employer code to a Business or Enterprise workspace or an API organization, which the email and OpenAI's training article place outside the change.
In the Codex app, leave the Daybreak toggle off for work that does not need it. OpenAI says requests then run under standard safeguards, and its data controls page says the training setting still governs activity outside Trusted Access sessions.
Individuals who call Daybreak Blue through the API should plan for November 20.
Watch OpenAI's data controls page for the opt-out before November 29; it is not there yet.
OpenAI has said it will explain how the opt-out works closer to November 29. For more on how AI companies handle training opt-outs, see our report on a shared opt-out signal.
