OpenAI plans to use cybersecurity session data for model training starting November 29. The company confirmed the change in an email to Trusted Access program members. Users who previously turned off the ‘improve the model’ setting will now have data used for training.
The decision follows user backlash and prompted OpenAI to promise a separate opt-out mechanism. The move revives concerns about the Buckmaster-Alpöge case involving the unauthorized use of unpublished research. Interactions flagged during safety reviews remain exempt from training and are reserved for safety purposes.
Researchers worry this change sets a precedent for using sensitive data without explicit consent. This policy shift affects users in the cybersecurity sector who rely on secure environments. Developers must review their privacy settings carefully to ensure their session data is not used for training.
