Security teams using OpenAI’s Daybreak Blue will have their sessions fed into model training from 29 November. The company told members of its Trusted Access programme by email, and the new rule applies even to people who had switched off the “improve the model for everyone” option.
That is a reversal. Until now its sessions sat outside training altogether, which mattered to people who paste in vulnerability details, logs and client system layouts. Complaints arrived fast, and a second email followed with a promise of a dedicated opt-out. Nothing we read says when it will appear or how it will be switched on.
Separately, conversations that trip a safety review remain usable for safety work. The Decoder reported the change and credits Yoav Alon as its source. We found no public OpenAI notice, so everything here comes from that account of the emails.
Why researchers are uneasy
The unease echoes an earlier row over Codex. Two mathematicians, Tristan Buckmaster and Levent Alpöge, gave it drafts on the Navier-Stokes equations; OpenAI then showed its own proof and said their inputs had not been used. Even so, the doubt stuck.
For anyone with a Daybreak Blue seat, the date to watch is 29 November, when the default changes unless the opt-out has arrived.
