Skip to content
Neural Daily

BusinessNews

OpenAI to train on Daybreak Blue session data from Nov 29

🇭🇺 Olvasd el magyarul →

Starting November 29, OpenAI will use data from Daybreak Blue cybersecurity sessions to train its models. That data was previously excluded, even for users who allowed training. After backlash, the company promised a separate opt-out.

By: Mikola Máté Patrik · · 2 min read

OpenAI documentation page on models and Trusted Access for cyber
Image: OpenAI · Screenshot · Screenshot of the official OpenAI cyber safety documentation · Source

What's changing

Daybreak is OpenAI's Trusted Access for Cyber program for verified security professionals. The Blue tier offers fewer refusals for defensive work, while Red unlocks cyber-specific models. An email to program members says Blue session data can enter training from November 29 (The Decoder).

That is a reversal: until now the data was excluded from training regardless of whether users had the "improve the model for everyone" setting on. Interactions flagged during safety reviews can still be used for safety purposes.

Why practitioners are uneasy

Blue sessions often involve vulnerabilities, internal code and details of client systems. After swift backlash, OpenAI sent a second email promising a separate opt-out, without details yet.

According to The Decoder, researchers fear a repeat of an earlier case. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes drafts to Codex, and OpenAI soon presented its own proof. The company denied training on their inputs, but the suspicion alone showed what is at stake.

What to do

If your team uses Daybreak Blue with client data, review your contract terms before November 29 and enable the opt-out as soon as it is available. Until then, assume anything you upload may become training data.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles