OpenAI plans to begin using Daybreak Blue cybersecurity session data to train its models starting November 29. The change, confirmed in an email to Trusted Access program members, reverses a previous exclusion of this data from training.
The announcement prompted swift backlash. OpenAI then sent a second email promising a separate opt-out, though interactions flagged during safety reviews can still be used for safety purposes.
A change to how session data is handled
Previously, Daybreak Blue session data was excluded from model training even when users had the “improve the model for everyone” setting turned off. The planned change means the existing setting will no longer, by itself, describe how this particular session data is treated.
The follow-up promise of a separate opt-out gives members another choice about training use. The source does not detail how that opt-out will work, when it will be available, or whether it changes the handling of data flagged in safety reviews.
That distinction matters: OpenAI says those flagged interactions may still be used for safety purposes. Training and safety review are described as separate uses, and the promised opt-out applies to the former rather than removing the latter.
Why researchers are paying attention
The shift has raised concern among researchers about how work shared with AI systems might be handled. The source connects those worries to the case involving mathematicians Tristan Buckmaster and Levent Alpöge.
They uploaded drafts of Navier-Stokes equations to OpenAI’s Codex. OpenAI soon presented its own proof, but denied using their inputs for training. The account does not establish that their drafts were used to produce the proof.
For researchers, the episode illustrates how a close sequence of events can trigger suspicion even when a company denies using submitted material for training. The source says the suspicion itself was enough to show researchers what could go wrong.
Trust depends on clear boundaries
When people share technical work with an AI service, they need to understand whether that material may be used to train models, reviewed for safety, or handled in another way. A setting called “improve the model for everyone” may not answer every question if a specific program’s data follows a different rule.
In this case, the original exclusion gave Daybreak Blue session data a distinct status, including when the general improvement setting was off. The announced change alters that status, and the promised separate opt-out is intended to give members a way to decline training use.
The safety-review exception remains relevant to that choice. Members may opt out of training use, according to the follow-up email, while interactions flagged during safety reviews can still be used for safety purposes. The source provides no further details about the criteria for flagging or the process for opting out.
What the announcement leaves open
The source establishes the planned start date, the previous exclusion, the backlash, and OpenAI’s promise of a separate opt-out. It does not specify what counts as a Daybreak Blue session, how long data is retained, or how members will exercise the new choice.
Those unanswered details shape how users can judge the change in practice. For members handling sensitive cybersecurity work, a clear account of the boundary between model training and safety review would help explain what the opt-out does and what uses remain possible.
For now, the stated policy shift is straightforward: OpenAI plans to train models on Daybreak Blue session data starting November 29, after previously excluding it even for users who had turned off “improve the model for everyone.” The separate opt-out addresses training, while safety-review use remains permitted.