OpenAI to train models on Daybreak Blue cybersecurity session data starting November 29

OpenAI will begin using Daybreak Blue session data to train its models starting November 29. The change was confirmed in an email to Trusted Access program members and applies to data that was previously excluded from training. The report says this includes cases where users have turned off general model-improvement training.

After swift criticism, OpenAI followed up with a second email promising a separate opt-out for Daybreak Blue session data. Users should watch for that dedicated opt-out rather than assuming existing training preferences cover the new policy. The report also notes that interactions flagged during safety reviews can still be used for safety purposes.

The change revives worries about whether researchers can safely use AI systems with unpublished work or sensitive technical material. The Decoder points to the Buckmaster-Alpöge case, in which mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to OpenAI’s Codex. OpenAI later presented its own proof and denied using their inputs for training, but the dispute became a warning about trust, consent, and perceived misuse.

Anyone using Daybreak Blue for sensitive cybersecurity work should review OpenAI’s upcoming opt-out process once available. The key issue is not only whether data is used for general training, but also how exceptions such as safety reviews are handled. For teams, this is a prompt to revisit internal rules for submitting confidential research, exploit analysis, or unpublished technical material to AI tools.