Starting November 29, OpenAI will use session data from Daybreak Blue users to train its models, a policy shift confirmed through an email sent to Trusted Access program members and reported by The Decoder. The update reverses a previous boundary where such data was excluded from training, even when users disabled the “improve the model for everyone” setting.
Policy Shift in Trusted Access Program
Daybreak functions as OpenAI’s Trusted Access for Cyber program, featuring security tiers such as Daybreak Blue and Daybreak Red. Vetted security researchers rely on these models—which feature fewer refusals in ChatGPT and Codex—for defensive tasks including vulnerability and malware analysis. According to details on OpenAI’s Help Center cited by aicoder.com, qualifying individuals must enable Advanced Account Security and register at least one physical FIDO2 hardware key.
Confusion? U-turn? pic.twitter.com/c1nNlnb4z3
Under the updated policy, OpenAI may use interaction content from Daybreak sessions to strengthen safeguards against cyber misuse, overriding individual training preferences. However, activity outside Trusted Access sessions continues to follow the user’s manual settings, and managed workspaces like Business, Enterprise, and Edu editions remain excluded from training by default, according to aicoder.com.
User Backlash and Promised Opt-Out
The announcement triggered immediate pushback within the security community. Security researcher Sean Heelan published a critical post on X drawing over 600 likes and 100 reposts, characterizing the move as taking researchers’ private work and predicting it would harm Daybreak usage, as noted by aicoder.com. Following the swift backlash, OpenAI issued a second email to program members promising a separate opt-out option.
Despite the promised opt-out, security researchers remain concerned about lingering ambiguities. Interactions flagged during safety reviews can still be used for safety purposes, leaving OpenAI with discretion over what qualifies as an exception.
Researchers have drawn parallels to past controversies, specifically the Buckmaster-Alpöge case where mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to Codex. Although OpenAI presented its own proof shortly afterward and denied using their inputs for training, the incident highlighted the inherent risks for researchers handling confidential data.
Frequently Asked Questions
What specific changes take effect on November 29?
OpenAI begins using Daybreak Blue session data to train its models, overriding individual opt-out settings for that specific program tier, according to emails sent to Trusted Access members and covered by The Decoder.
How does this policy affect API access and personal accounts?
User-shared emails posted on X indicate that individuals may lose Daybreak Blue API access on November 20, while Codex access via ChatGPT sign-in remains available. Managed workspaces and API platforms outside the personal Daybreak tier are not used for training by default.
What can developers and researchers do to protect sensitive data?
Developers using Daybreak Blue should avoid pasting unreported vulnerabilities, client asset information, or unsubmitted reports. Teams requiring stricter data boundaries can explore organization-level Daybreak access through managed workspaces, as detailed by aicoder.com.