OpenAI will start training on Daybreak Blue cybersecurity session data, then adds a separate opt-out after backlash

OpenAI will begin using Daybreak Blue session data to train its models starting November 29, according to an email sent to members of its Trusted Access program and reported by The Decoder. Until now, this cybersecurity session data was excluded from training, even for users who had turned off the "improve the model for everyone" setting. The Decoder says the backlash was swift, and OpenAI sent a second email promising a separate opt-out. Interactions flagged during safety reviews can still be used for safety purposes. The change has worried researchers, who point to the case of mathematicians Tristan Buckmaster and Levent Alpöge. The two uploaded drafts on the Navier-Stokes equations to OpenAI's Codex. OpenAI soon presented its own proof but denied using their inputs for training. The Decoder says the suspicion alone showed researchers what could go wrong. Those affected are security professionals and researchers who use Daybreak Blue under Trusted Access, and the organizations whose sensitive security work may pass through those sessions. The change matters because Trusted Access users may have chosen the program partly because their sessions were kept out of training. Once the new policy takes effect, keeping that data out of training will depend on users actively opting out.