OpenAI says Daybreak Blue sessions on personal accounts may be used for training even with the training toggle off, drawing security researcher backlash on X

Background and the problem

Daybreak is OpenAI's Trusted Access for Cyber program, with tiers such as Daybreak Blue and Daybreak Red. Vetted security researchers get models with fewer refusals in ChatGPT and Codex for defensive work such as vulnerability and malware analysis. According to the OpenAI Help Center, individuals must enable Advanced Account Security and register at least one physical FIDO2 hardware key to qualify. The catch is that researchers often feed the model unreported vulnerabilities, target details and proofs of concept, so whether that content ends up in training data is a hard line for them.

Architecture and how it works

OpenAI's updated "Data controls in ChatGPT" page adds a "Trusted Access on personal accounts" section. During Daybreak sessions, OpenAI may use interaction content to improve its models even when "Improve the model for everyone" is off, to strengthen safeguards against cyber misuse. Activity outside Trusted Access sessions still follows the user's own setting. Meanwhile, the Advanced Account Security page still says conversations are not used to train OpenAI models while that setting is enabled, and OpenAI has not explained how the two documents fit together. Managed workspaces (Business, Enterprise, Edu) and the API Platform are still not used for training by default.

Benchmarks and measured results

This is a data policy story with no model scores involved, and OpenAI has published no related evaluations. The official pages do not give an effective date. Emails that several users posted on X say the change starts November 29, and some users also say individuals lose Daybreak Blue API access on November 20 while Codex via ChatGPT sign-in stays; both points come from user-shared emails. On the community side, security researcher Sean Heelan's critical post drew more than 600 likes and 100 reposts, calling it taking researchers' private work and predicting it would hurt Daybreak usage.

Practical guide for developers

Individuals using Daybreak Blue should assume anything in a Daybreak session may be used for training, and avoid pasting details of unreported vulnerabilities, client asset information or unsubmitted reports. Teams that need a stricter data boundary can look at organization-level Daybreak access: OpenAI says managed workspaces and the API are not used for training by default, and organization access has its own onboarding process. Check your training setting under Settings > Data controls in ChatGPT, and watch the Help Center pages for an official effective date and scope.