OpenAI Starts Training on Daybreak Blue Data

OpenAI told members of its Trusted Access program this week that starting November 29, it will begin training models on Daybreak Blue cybersecurity session data. Until now, that data sat outside the training pipeline entirely, even for users who had switched off the "improve the model for everyone" setting. The policy flip landed in members' inboxes by email, and the reaction was fast enough that OpenAI sent a second message within days promising a separate opt-out mechanism.

The company says interactions flagged during safety reviews remain fair game for safety purposes regardless of what users choose. But the reversal has set off alarm bells among researchers who remember what happened with mathematicians Tristan Buckmaster and Levent Alpöge. The two had uploaded draft work on the Navier-Stokes equations to OpenAI's Codex, only to watch OpenAI later present its own proof on the same problem.

OpenAI denied training on their inputs, and nothing was ever proven either way. That unresolved doubt is exactly the problem now resurfacing: once trust in a data boundary breaks, no amount of denial fully puts it back together.