OpenAI to train on Daybreak Blue security sessions, then promises an opt-out

OpenAI planned to train its models on Daybreak Blue session data from personal accounts starting November 29, even for users who had switched off "Improve the model for everyone," according to The Decoder. Until now, that data was excluded from training.

OpenAI's data controls help page, updated October 30, says it "may use the content of your interactions to improve our models even when ‘Improve the model for everyone’ is turned off," Neural Brew reported. The toggle still governs activity outside Daybreak. Business, Enterprise, Edu and ChatGPT for Healthcare workspaces, plus API data, stay untrained by default.

Backlash was swift. OpenAI sent a second email promising a separate opt-out for personal accounts, with more detail closer to November 29, RuntimeWire reported. Neural Brew reported that the help page itself mentions no opt-out. The same RuntimeWire piece notes API access for individuals ends November 20.

The context: Daybreak is OpenAI's Trusted Access for Cyber program, and Blue covers incident response and malware analysis. RuntimeWire says sessions can contain unpatched vulnerabilities, exploit proofs of concept and incident data. OpenAI says it needs these interactions to bolster defenses against malicious use of AI.

My take: an opt-out that exists only as a promise is not a control, it is a press release with a toggle-shaped hole. Mechanism, retention period and human review are all unanswered.