OpenAI to train on Daybreak Blue data unless users opt out

OpenAI plans to start using Daybreak Blue cybersecurity session data to train its models on November 29, according to The Decoder. The change matters because the same data had previously been treated as off-limits, and users will need a separate opt-out to keep it out of training.

The Decoder reports that OpenAI confirmed the change in an email to members of its Trusted Access program. After quick criticism, the company reportedly sent a follow-up message saying users would get a dedicated way to opt out.

What is changing on November 29

According to The Decoder, Daybreak Blue session data will become eligible for model training starting November 29. The outlet says this is a shift from the prior handling of that data, which had not been used for training even in relation to the broader “improve the model for everyone” setting.

The important practical detail is that the new opt-out is separate. In other words, users in the affected program should not assume that a general training preference is enough to cover Daybreak Blue sessions. The Decoder says OpenAI promised a separate opt-out after backlash to the initial email.

There is also a safety carve-out. The report says interactions flagged during safety reviews may still be used for safety purposes, even as OpenAI adds an opt-out for training use.

For anyone using OpenAI products for sensitive work, including tools such as ChatGPT, the lesson is straightforward: data settings can be product- or program-specific, and they can change. Teams should treat any new email about model improvement, safety review, or training data as something that needs review rather than a routine policy notice.

Why the research community is concerned

The Decoder connects the reaction to an earlier controversy involving mathematicians Tristan Buckmaster and Levent Alpöge. According to the report, the pair uploaded drafts related to the Navier-Stokes equations to OpenAI’s Codex. OpenAI later presented its own proof and denied that their inputs had been used for training.

The point, as The Decoder frames it, is not only whether training happened in that case. The suspicion itself showed how fragile trust can be when people submit unfinished, unpublished, or high-value work to AI systems.

That concern is especially sharp for cybersecurity sessions. The source does not specify what Daybreak Blue sessions contain, but the label alone makes the consent question more sensitive for researchers and security teams. If a session includes tactics, vulnerabilities, analysis, or proprietary defensive work, users may want clearer boundaries around reuse.

What affected users should check now

The report says the notice went to Trusted Access program members, so those users are the most directly affected based on the available information. They should look for the follow-up message from OpenAI and confirm how the promised separate opt-out works before November 29.

A useful internal checklist would be:

• Identify who in the organization receives OpenAI program emails.
• Confirm whether Daybreak Blue is being used for sensitive research or client work.
• Check whether a separate opt-out has been offered and recorded.
• Review what “safety purposes” means for workflows that may be flagged.
• Tell team members not to assume older data settings still cover all sessions.

For freelancers, consultants, and small security teams, the broader issue is client trust. If you use AI tools while handling confidential material, clients may reasonably ask whether their data could be used to improve a vendor’s models. A separate opt-out requirement makes documentation more important.

A wider signal on AI data consent

This story fits into a larger pattern: AI companies are under pressure to make training permissions more explicit. Emonarc recently covered how OpenAI and other major AI companies backed a shared AI scraping opt-out, another sign that consent mechanisms are becoming a competitive and legal flashpoint. See: OpenAI, Google, Microsoft and Anthropic back shared AI scraping opt-out.

It also lands in a period when privacy controls inside AI products are getting more attention, including consumer-facing features such as ChatGPT locked chats. Those tools address a different layer of privacy, but they reinforce the same takeaway: users need to understand both who can access their conversations and whether those conversations can train future models.

For OpenAI, the near-term question is whether the separate opt-out is clear enough to rebuild confidence among Daybreak Blue users. For everyone else, the safer habit is to assume that sensitive AI workflows need periodic policy checks, not a one-time settings review.