OpenAI has told security researchers in its Trusted Access programme that it will start using their Daybreak Blue session data to train its models from November 29, even when they have switched training off. The change reverses a previous rule under which Daybreak Blue sessions were excluded from training entirely, and it has drawn a sharp backlash from the very researchers the programme was built to serve.
Daybreak Blue is the lower tier of OpenAI's Trusted Access for Cyber programme, which gives vetted security researchers access to its models with fewer restrictions for defensive work such as vulnerability triage, secure code review and incident response. That work routinely involves unpublished vulnerabilities, exploit proofs of concept and details from live investigations — material researchers cannot simply hand over as training data.
According to OpenAI's updated data-controls help page, during Daybreak sessions on a personal account, the company may use content from interactions to improve its models even when the "Improve the model for everyone" setting is turned off. The setting continues to govern a user's activity outside Trusted Access sessions, and Business, Enterprise, Edu and API Platform services remain excluded from training by default.
A reversal researchers spotted first
The policy shift came to light through emails sent to Trusted Access members, first reported by The Decoder. Security researcher Sean Heelan flagged one of the emails on X, noting that the change applied even to users who had explicitly turned training off, in a post that drew hundreds of likes and reposts across the security community.
The reaction was swift. Researchers argued that sessions built for defensive work could contain unpatched flaws and exploit chains that, once absorbed into a model, could surface in other contexts. OpenAI's stated reason for the change is to strengthen safeguards against cyber misuse, but critics questioned why that goal required overriding an explicit opt-out.
Within hours, OpenAI sent a second email to programme members promising a separate opt-out for Daybreak sessions. That control is not yet described on the company's public help page, and it remains unclear when it will appear, what its default will be, or whether it will apply to past sessions.
The opt-out comes with a catch
Even the promised opt-out may not cover everything. OpenAI has said that interactions flagged during safety reviews can still be used for safety purposes, leaving the company discretion over what qualifies as an exception. For researchers, that distinction matters: opting out of model training would not necessarily prevent all use of session data.
User-shared emails also indicate that individual users may lose Daybreak Blue API access on November 20, a step that would push researchers who need programmatic access toward business accounts, which keep their existing training protections.
What stays protected — and what does not
The change applies only to Daybreak sessions on personal accounts. Activity outside Trusted Access sessions remains governed by the user's own training setting, and managed workspaces such as Business, Enterprise and Edu editions are not used for training by default.
Still, the episode has shaken confidence in a programme whose entire premise was trusted access. Researchers joined Daybreak to test defences against real threats; the fear now is that the work they do inside it could end up training the next generation of models, including models available to the attackers they are trying to stop. OpenAI has said more detail will come closer to November 29, but for now, the researchers most affected are waiting to see whether the promised opt-out arrives before the training begins.



