The story behind the cartoon
Starting November 29, OpenAI will use Daybreak Blue session data to train its models, according to an email to members of its Trusted Access program reported by The Decoder. The sessions had previously been excluded from training, even for users who had switched off the “improve the model for everyone” setting.
Daybreak Blue gives vetted security researchers access to OpenAI's models for defensive cyber work — sessions that can hold unpatched vulnerabilities, exploit proofs, and incident details.
After swift backlash, OpenAI sent a second email promising a separate opt-out, though interactions flagged in safety reviews can still be used for safety purposes; the change revives researchers' worries from the Buckmaster–Alpöge dispute, in which mathematicians suspected unpublished drafts they uploaded to Codex had fed OpenAI's own work, which OpenAI denied.
The researcher flipped the big TRAINING switch to OFF with great ceremony, while the robot beside him quietly ate his briefcase — there was, it turns out, a second, much smaller switch, and it was already on.

