OpenAI Will Train on Daybreak Blue Security Sessions From November 29

OpenAI plans to start training its models on data from Daybreak Blue sessions on November 29. Daybreak Blue is a cybersecurity offering for members of OpenAI’s Trusted Access programme, a vetted tier for security professionals. Until now, this data was off limits for training, even when a user had switched off the “improve the model for everyone” setting.

Members learned of the change through an email, according to The Decoder and a post shared on LinkedIn. After backlash, a second email promised a separate opt-out for this particular use. How that opt-out works and when it arrives has not been explained. Interactions flagged during safety reviews can still be used for safety purposes, which is standard practice across AI providers.

Why would this upset people? Security sessions can contain sensitive material: details of a company’s systems, suspected vulnerabilities, or code under review. Professionals often work under confidentiality agreements, so “your data may now help train a model” lands differently for them than for someone asking for a recipe. The Decoder also links the story to a recent dispute in which two mathematicians, Tristan Buckmaster and Levent Alpöge, uploaded unfinished Navier-Stokes proof drafts to OpenAI’s Codex. OpenAI later presented its own proof and denied training on their inputs. Researchers say the new policy revives the same trust question.

What’s behind it: Training data is the fuel of AI models, and conversations with experts are valuable fuel. Companies typically offer a switch, and the fair question is whether the switch means what it says. When an exclusion that users counted on is removed for one product, trust takes the hit even if an opt-out follows. It is also worth being precise about what we know: the details come from an email forwarded by recipients, not from a public OpenAI policy page, so the final wording may differ.

What this means for you: If you are a Trusted Access member, check your account settings and the follow-up email for the opt-out before November 29, and avoid pasting client material into any session until you know where it ends up. If you are an ordinary ChatGPT user, this change does not apply to you directly, but it is a good moment to look at your own data controls and ask a simple question of every AI tool at work: is what I type used for training, and can I turn that off in writing?

Sources

• The Decoder: OpenAI plans to train models on Daybreak Blue cybersecurity session data
• Backlash post on X