OpenAI promises a separate Daybreak training opt-out after notice of an exception

OpenAI says it may use Daybreak Blue sessions on personal accounts to improve its models even when the ordinary training setting is off. Its current data-controls help page describes an exception intended to strengthen safeguards against cyber misuse. The existing training preference continues to govern activity outside Trusted Access sessions.

An initial email, reported received on October 30 and publicly shared by recipients, gives November 29 as the start date. A follow-up email says users will have a separate option to opt out after that date. Interactions flagged for safety review may still be used for safety purposes, including improving protections against harmful use.

The change affects researchers enrolled through the personal-account route. That route requires a paid plan, Advanced Account Security and a compatible physical security key. OpenAI describes Blue as access to supported general-purpose models with fewer unnecessary refusals for authorized defensive work, including code review, malware analysis and patch validation.

The current Advanced Account Security help page says conversations are excluded from model training while the feature is enabled. Its account-takeover protections include passkey or security-key sign-in, with passwords and email or SMS sign-in codes disabled.

The program limits access to approved internal security work on systems the user owns or has explicit permission to test. Its overview ties access to approved workspaces or API projects. Customer-facing services and extending access to external users fall outside that scope.

Existing terms for enterprise safety processing

The initial notice preserves existing treatment for Business and Enterprise customers and for ordinary Advanced Account Security conversations. OpenAI’s enterprise privacy commitments keep business data out of training by default. Explicitly shared data, such as through opt-in feedback, may be used for training under the company’s stated policy.

The initial email also preserves existing treatment for Private Safety Processing and Zero Data Retention. The PSP guide describes automated safety review without OpenAI retaining customer prompts or responses. Encrypted records reside in customer-controlled storage and are reviewed inside an attested runtime that disables human access. Only bounded safety signals and operational metadata leave that protected review in plaintext. Content in customer storage serves approved safety purposes and is excluded from model training.

The API data guide separates training use from abuse-monitoring retention. Logs may contain prompts, responses and classifier metadata. Default retention is up to 30 days, with extensions for legal requirements or harm prevention. Zero Data Retention and Modified Abuse Monitoring require prior approval. Some features also persist application state, so storage depends on the API capability used.

In the public developer forum, a recipient who reported receiving the notice on October 30 asked about an explicit opt-out and whether leaving the program would be necessary. Another participant asked about earlier sessions, uploaded files and context obtained through tools. These are requests for clarification, not evidence that historical sessions or every category of uploaded material will be used for training.

OpenAI’s follow-up says it will explain how the separate choice works closer to November 29.