OpenAI plans to use data from Daybreak Blue cybersecurity sessions to train its models starting November 29. An email to members of the Trusted Access program confirmed the change, according to The Decoder. The sessions had previously been excluded from training, even for users who had turned off the general “improve the model for everyone” setting.
After members pushed back, OpenAI sent a second email promising a separate opt-out for Daybreak Blue sessions. The emails do not establish how that opt-out will work or whether it has to be enabled by each user. OpenAI says interactions flagged during safety reviews may still be used for safety purposes, so opting out of model training would not necessarily prevent all use of session data.
Why researchers are concerned
The policy shift raises concerns about whether material shared in specialized research or technical sessions might later contribute to model training. The Decoder points to mathematicians Tristan Buckmaster and Levent Alpöge, who uploaded drafts about the Navier-Stokes equations to OpenAI’s Codex. OpenAI later presented its own proof and denied using their inputs for training.
That denial did not dispel the researchers’ suspicion, and the episode illustrated the risks that can arise when people share unpublished work with an AI tool. The account does not establish that OpenAI used the mathematicians’ drafts to train a model. It is a separate example of why researchers may care about clear limits on how submitted material is handled.
What users should know
For Trusted Access members, the important change is that Daybreak Blue sessions are no longer covered by the previous blanket exclusion from training. The general setting to improve models for everyone will not, by itself, preserve that exclusion under the announced plan.
OpenAI has promised a separate way to opt out, but the available details do not say when it will appear, what its default will be, or whether it applies to past sessions. The planned start date is November 29; the supplied announcement does not specify a year. Safety-review use is also a separate exception from training, so the two purposes should not be treated as the same setting.
Why it matters
Trusted Access members who use Daybreak Blue should not assume that turning off the general model-improvement setting keeps those sessions out of training. OpenAI has promised a separate opt-out, but its timing and default are not specified; safety-review use may still continue.
