OpenAI will start training on security researchers' Daybreak Blue sessions, and the opt-out does not cover the most sensitive ones.
AspireAI community · The Decoder
**OpenAI will start training on security researchers' Daybreak Blue sessions, and the opt-out does not cover the most sensitive ones.**
Daybreak Blue is the recommended tier of OpenAI's Trusted Access for Cyber program, a gated track for authorised defensive security work: secure code review, vulnerability triage, malware analysis, incident response. Individual access needs Advanced Account Security and a FIDO2 key, and enrolling turns the "improve the model for everyone" setting off.
On October 30 OpenAI emailed those members saying that from November 29 it may use Daybreak Blue sessions on personal accounts to improve its models, even with that setting off. Previously this data was excluded from training entirely. OpenAI's data controls page now carries the same rule with no date on it. Backlash was swift, and a second email promised a separate opt-out. It also says interactions flagged for safety review may still be used for safety purposes. Business and Enterprise customers are unaffected.
The honest limit: the promised opt-out was not on the help page when reporters checked, and OpenAI has not said how a session gets flagged or who reviews it. The sessions most likely to trip a safeguard are the ones the program exists for.
https://the-decoder.com/openai-plans-to-train-models-on-daybreak-blue-cybersecurity-session-data-that-was-previously-off-limits/
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.