Key points
- Starting November 29, OpenAI will train its models on session data from Daybreak Blue, its cybersecurity offering, which was previously kept out of training.
- The earlier exclusion held even for users who had switched off the setting that lets their data improve models for everyone.
- After complaints, OpenAI sent a second email to Trusted Access members saying a separate opt-out will be offered.
- Researchers cite a dispute in which two mathematicians suspected OpenAI of using their uploaded drafts; OpenAI denied training on them.
Why it matters for marketers: Teams that put sensitive client or security material into AI tools should recheck data-use terms and opt-out options, because a policy default can change with one email.
OpenAI told members of its Trusted Access program that, beginning November 29, it will use Daybreak Blue session data to train its models. Until now, that data was kept out of training, and the exclusion applied even when a user had turned off the setting that lets their data help improve models for everyone, according to The Decoder.
The reaction was fast. OpenAI sent a second email to members saying a separate opt-out will be available. The report adds that OpenAI can still use interactions that are flagged in safety reviews for safety work.
The change revived a worry among researchers. Two mathematicians, Tristan Buckmaster and Levent Alpöge, had shared Navier-Stokes equation drafts with OpenAI's Codex. OpenAI later showed its own proof but said it had not trained on their inputs. That denial did not end the doubt, and the episode showed how suspicion alone can erode trust.
For marketers, agencies and content teams, the lesson reaches beyond security work. Campaign plans, unreleased copy, client data and strategy documents often pass through AI tools under the assumption that a privacy setting settles the matter. This case shows that a default can shift, and that a toggle you relied on may not cover a new category of use.
Next steps are practical. Watch for how the promised opt-out works and when it arrives, and read the data terms for each AI tool your team uses. If you handle confidential client material, check whether protections are contractual or only a setting that can change.
Sources
This brief is our own summary. Read the full reporting at the source.
