OpenAI plans to start using Daybreak Blue cybersecurity session data to train its models beginning November 29, reversing a previous policy that kept this information off-limits. The update was communicated to Trusted Access program members via email, noting that this data was previously excluded from training even when users enabled the model improvement setting. Following swift backlash from users, OpenAI issued a secondary email to promise a separate opt-out mechanism for participants, while clarifying that interactions flagged during safety reviews remain eligible for safety training purposes.
The policy shift has raised concerns among researchers who worry about a potential repeat of the Buckmaster-Alpöge case. In that previous situation, mathematicians Tristan Buckmaster and Levent Alpöge uploaded unpublished drafts regarding the Navier-Stokes equations to OpenAI's Codex. Although OpenAI subsequently presented its own proof and denied using their specific inputs for model training, the episode left researchers wary of potential intellectual property risks.
This development highlights ongoing tensions between AI developers seeking vast data pools for model training and users concerned about the privacy of sensitive inputs.