OpenAI plans to train on cybersecurity session data, drawing researcher protests
THE DECODER reports that OpenAI will train its models on cybersecurity session data, and researchers have protested the plan.
OpenAI plans to start training its models on data from Daybreak Blue sessions beginning November 29, reversing an earlier policy under which that data was excluded from training. According to THE DECODER, citing an email sent to Trusted Access program members, the data had been excluded from training even for users who had the "improve the model for everyone" setting turned off. Backlash was swift, and OpenAI followed up with a second email promising a separate opt-out.
Daybreak Blue gives vetted security researchers access to models with fewer restrictions on security work. Its sessions can involve unpublished vulnerabilities, exploit chains and details from ongoing investigations. Sean Heelan, credited in a LinkedIn post, flagged the email telling individual users that the change starts November 29. The post notes that OpenAI's data controls page already says it may use Daybreak sessions on personal accounts to improve its models regardless of the training setting. Business and enterprise accounts keep their existing protections, while individual researchers get no opt-out for Daybreak, according to the same post.
Interactions flagged during safety reviews can still be used for safety purposes, the report says.
THE DECODER links the backlash to an earlier episode. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to OpenAI's Codex. OpenAI soon presented its own proof but denied using their inputs for training. The report says the suspicion alone was enough to show researchers what could go wrong.
In a post on X, John Kelleher called the change outrageous. He described it as applying to Daybreak Blue accounts with Advanced Account Security, which would share data for training even when that setting is disabled.
What changed
OpenAI will start training its models on Daybreak Blue session data, reversing an earlier policy under which that data was excluded from training, including for users with the training setting turned off.
Type: Breaking change · Applies to: Daybreak Blue
Why it matters
Companies that share security session data with AI vendors should check how that data may be used for training.
What you can do
Review contracts and data-sharing settings before sending security session data to AI vendors.
Original source
Further reporting
Where we saw it
Matthias from THE DECODER
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works