THE ESSENTIALS IN 30 SECONDS

  • According to emails reported by The Decoder and shared by users, OpenAI will use Daybreak Blue session content from personal accounts for model training starting November 29.
  • The change reportedly applies even when the 'improve the model for everyone' setting is off; this data had previously been excluded from training.
  • After a backlash, a second email promised a separate opt-out, but its details and timing have not been published.

OpenAI has told members of its Trusted Access for Cyber program, branded Daybreak, that it will start using content from Daybreak Blue sessions to train its models on November 29. The change was disclosed by email rather than in a public announcement, and it quickly drew criticism from security researchers who use the service.

What the emails say

According to The Decoder, which saw the message, the change applies even to users who have switched off the "improve the model for everyone" setting. Until now, Daybreak Blue sessions had been kept out of training. A user who posted the email on OpenAI's developer forum on October 30 said it concerned Daybreak Blue sessions on personal accounts. In a blog post, the AI company CellCog, which also reviewed the email, reports that business and enterprise customers and Zero Data Retention setups are not affected. Interactions flagged during safety reviews can still be used for safety purposes, The Decoder notes.

After the backlash, OpenAI sent a second email promising a separate opt-out, The Decoder reports. Neither the publication nor other reports explain how that opt-out will work, when it will be available, or whether it will cover data collected before it exists. OpenAI has not commented publicly beyond the emails, and its help pages that we could open do not yet set out the new terms.

What Daybreak Blue is

Daybreak is OpenAI's vetted access program for cyber defenders. Its help pages describe two levels: Daybreak Blue, which reduces refusals for security work on supported models via the API and Codex, and Daybreak Red, which requires separate approval and gives access to cyber-specialised models. Individual users had to enrol in Advanced Account Security and register a physical security key by October 1. OpenAI also stresses that some safeguards and usage policies remain in place.

Why researchers object

Security work often involves sensitive material such as unpatched vulnerabilities, client systems or exploit techniques. Forum posters said they would stop using Daybreak Blue without a way to keep their data out of training, and one questioned whether earlier sessions could be affected. The Decoder links the reaction to a recent dispute in which mathematicians suspected their Codex inputs had been reused, which OpenAI denied. Until the opt-out is published, the exact scope of the change rests on the emails as reported.