OpenAI to fold Daybreak Blue security sessions into model training

From November 29, OpenAI will train on Daybreak Blue session data, which had been excluded even for users who opted out. A separate opt-out is promised.

OpenAI plans to begin training its models on Daybreak Blue session data on November 29. It told members of its Trusted Access program by email. Until now, that data was kept out of training, including for users who had switched off the "improve the model for everyone" setting.

The change drew quick backlash. OpenAI then sent a second email promising a separate opt-out. Interactions flagged during safety reviews can still be used for safety purposes, so some use of the data continues whatever a member chooses.

This matters for security because defensive work involves sensitive material: vulnerabilities, internal systems, and incident details. Teams will use AI tools for that work only if they trust where the data goes. Removing a protection that members had relied on makes the tools harder to adopt for the people who most need to defend themselves. It also adds friction to the wider spread of AI-assisted security. Researchers point to the Buckmaster-Alpöge case as a similar worry. In that case, mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to OpenAI's Codex. OpenAI later presented its own proof but denied using their inputs for training.

Watch for the details of the promised opt-out, including when it arrives and whether it restores the earlier exclusion. Several caveats apply. The article gives no timing or specifics for that opt-out. The mathematicians' case is only suspicion, since OpenAI denied using their inputs. The source is a short news item with limited detail on Daybreak Blue and Trusted Access, so the full reporting at The Decoder is worth reading.

Source: The Decoder