Daybreak Blue gives vetted security researchers access to models with fewer restrictions, covering unpublished vulnerabilities, exploit chains, and details from ongoing investigations.

Sean Heelan flagged an email to Trusted Access program members confirming the shift. Previously, the data was excluded from training regardless of the setting.

Researchers pushed back quickly. OpenAI followed with a second email promising a separate opt-out, though the document does not confirm that the opt-out is technical.

Business and enterprise accounts keep their existing protections. The change affects personal accounts only.

The Buckmaster-Alpöge case is the warning. Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes drafts to Codex, and OpenAI presented its own proof while denying it used their inputs. The suspicion itself was enough to show what could go wrong.

The exact corpus size, the scope of safety-review reuse, and whether the opt-out can be verified in practice remain unmeasured. Flagged interactions and ordinary training data are not yet separated in the document.

The date and scope are confirmed: personal accounts lose the training exclusion on November 29 while enterprise protections hold. The opt-out is a promise, not a verified mechanism, and the Buckmaster-Alpöge precedent explains why researchers will be skeptical.