OpenAI to train models on classified cybersecurity sessions from November 29

Images

×

OpenAI's email to Trusted Access members about the policy change for Daybreak Blue sessions. the-decoder.com

Starting November 29, OpenAI will begin using cybersecurity session data from its Daybreak Blue program to train its models — data the company previously excluded from training even when users had opted in to sharing. An email to members of OpenAI's Trusted Access program, who test models in private, high-stakes sessions on classified and pre-release systems, confirmed the reversal. After swift backlash, a second email promised a separate opt-out for Daybreak Blue data.

The change matters because of who uses these sessions. Daybreak Blue is the company's channel for US government and critical-infrastructure security work: researchers and officials probing frontier models for vulnerabilities in nuclear plants, pipelines and defence systems. That material — exploit drafts, threat intelligence, the private tradecraft of state defenders — now feeds the training corpus unless each participant notices a policy email and opts out. OpenAI notes that interactions flagged during safety reviews can still be used for safety purposes, a carve-out that keeps the company's own discretion intact while the participants' shrinks.

Researchers immediately raised the Buckmaster-Alpöge precedent. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded their drafts on the Navier-Stokes equations to OpenAI's Codex; shortly afterward, OpenAI presented its own related proof. OpenAI denied training on the mathematicians' inputs, but no one outside the company can audit the denial. The suspicion alone settled the question of what a private repository of expert work is worth once the operator's business model is the corpus itself.

The arithmetic is simple enough. Government threat researchers gain early access to models worth millions in compute; OpenAI gains session logs from the most skilled security practitioners alive. Both sides signed up believing the data stayed put. As of November 29, one side changed the terms by email.