OpenAI Will Train on Daybreak Blue Cybersecurity Session Data From November 29, Reversing a No-Training Guarantee

OpenAI will start using session data from Daybreak Blue, the cybersecurity tier of its Trusted Access program, to train its models on November 29, according to an email to program members first reported by The Decoder on October 31. The change reverses a guarantee that had held regardless of user settings: until now, Daybreak Blue sessions sat outside the training pipeline entirely, even for members who had turned off the "Improve the model for everyone" switch on their accounts.

Daybreak Blue is the entry tier of Trusted Access for Cyber, which gives vetted security researchers reduced refusals on defensive work such as vulnerability triage, secure code review and incident response. Those sessions routinely contain exactly the material researchers cannot afford to expose: unpublished vulnerability details, exploit chains, server logs and the layout of client systems. Under the new policy, that content from personal accounts becomes training-eligible by default.

OpenAI's updated data controls page now makes the exception explicit: the company may use content from Daybreak Blue cybersecurity sessions on personal accounts to improve its models even when the account's master training switch is off. Activity outside those sessions still follows the user's own setting, and inputs and outputs from Business, Enterprise, Edu and API Platform accounts are not used for training by default.

The backlash was fast. Security researcher Sean Heelan published a critical post on X that drew more than 600 likes and around 100 reposts, arguing that OpenAI was effectively taking researchers' private work and predicting the change would reduce Daybreak usage. Within days, OpenAI sent a second email promising a separate opt-out mechanism for Daybreak sessions. What has not been published is the detail that matters: when the control will appear, whether it will be on or off by default, and whether it covers sessions that occurred before November 29.

Researchers are also recalling an earlier dispute that has nothing to do with cybersecurity but shapes how the policy lands. Mathematicians Tristan Buckmaster and Levent Alpoge uploaded drafts of Navier-Stokes equation work to OpenAI's Codex; OpenAI later presented its own proof of a related result and denied using their inputs for training. The denial did not dispel the suspicion, and program members cited the episode when the November 29 change spread, as The Decoder noted.

One carve-out survives the change: interactions flagged during safety reviews can still be used for safety purposes, so opting out of training would not necessarily prevent all use of session data. OpenAI has said the training use is intended to strengthen safeguards against cyber misuse.

The practical consequence is a trust reset on a fixed date. Members of a program built on discretion now have to decide, before November 29, whether to keep pasting live exploit code and client vulnerability data into Daybreak Blue while waiting for an opt-out that exists only as a promise in an email. If OpenAI does not publish the control and its default state on a public policy page in the next four weeks, the likely outcome is exactly what Heelan predicted: sensitive sessions move somewhere else, and the program's main asset, researcher trust, erodes with them.