Skip to content
NewsProduct launch

OpenAI will train on Daybreak Blue cybersecurity session data it once excluded entirely

· by Pondero Newsdesk

The short version

OpenAI told vetted Trusted Access for Cyber members by email that Daybreak Blue penetration-test and malware-analysis sessions become training-eligible on November 29, reversing a no-training guarantee that had held regardless of a member's own data settings.

OpenAI will train on Daybreak Blue cybersecurity session data it once excluded entirely

OpenAI told members of its vetted Trusted Access for Cyber program, by email, that starting November 29, 2026, it will begin training its models on Daybreak Blue session data, per OpenAI's email reported by The Decoder. That session data, the live penetration-test, vulnerability-triage, and malware-analysis work Trusted Access members generate, had previously been excluded from training entirely, no matter what a member's own data settings said.

What changed, and what did not

Daybreak Blue is the entry tier of OpenAI's Trusted Access for Cyber program, which gives vetted enterprise security teams and researchers reduced refusals on mainline models, including GPT-5.6 Sol and GPT-6 Sol, for authorized defensive work such as vulnerability triage, detection engineering, and incident response. Per the email reported by The Decoder, Daybreak Blue session data moves on November 29 from a hard training exclusion to training-eligible data, reversing a guarantee that previously held even when a member had the "improve the model for everyone" setting turned off. Backlash from program members followed quickly, and OpenAI sent a second email promising a separate opt-out mechanism, per the same report. Interactions already flagged during safety reviews stay usable for safety purposes, unchanged from the prior policy. As of this writing, OpenAI's public Daybreak help pages, including the troubleshooting article covering data retention and Zero Data Retention settings, make no mention of the November 29 change; the only documentation of it is the vendor email, described secondhand.

Researchers have until November 29, with no opt-out mechanism published yet

A Trusted Access session routinely carries live exploit code, malware samples, and client vulnerability data that members previously pasted into Daybreak Blue trusting it would never train the model. That trust question now resets on a fixed date, and the opt-out OpenAI promised has not shipped. Members who want to keep that guarantee have to either stop feeding sensitive material into Daybreak Blue before November 29 or wait on a mechanism OpenAI has described only in a follow-up email, not a public policy page. The reversal lands in a program whose members were already primed to distrust OpenAI's training claims: per The Decoder, mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation proof drafts to OpenAI's Codex, OpenAI later presented its own version of a related proof, and OpenAI denied training on their inputs. That dispute has nothing to do with Daybreak, but Trusted Access members cited it when news of the November 29 change spread, per the same report.

What to watch next

Whether OpenAI publishes the promised opt-out mechanism, and the policy itself, on a public Daybreak page before November 29, and whether named Trusted Access partners comment on the change before the effective date arrives.

Sources