What changed
OpenAI plans to train its models on Daybreak Blue cybersecurity session data starting November 29. This data was previously excluded from training, even when users had the 'improve the model for everyone' setting disabled. The change was communicated to Trusted Access program members via an email.
What this means in practice
The decision to include Daybreak Blue data in training comes after a backlash against the initial announcement. OpenAI responded with a second email promising a separate opt-out option. However, interactions flagged during safety reviews can still be used for safety purposes. Researchers are concerned about potential misuse, citing the Buckmaster-Alpöge case as a precedent.
Limitations and availability
The change has not been made available to the general public yet. Only Trusted Access program members were informed of the policy change. The data used for training will be subject to the same safety review processes as other data. The supplied announcement does not specify the exact scope of the data or the extent of the opt-out mechanism.
