OpenAI says it’s tightening the locks on its models’ private reasoning
OpenAI reports a 0.2% success rate for extracting earlier reasoning in an internal test of updated safeguards.
TLDR
Distillation trains one model on another’s outputs. OpenAI says its defenses combine tighter protections for private reasoning with investigations into coordinated extraction campaigns. In an internal evaluation, earlier-reasoning extraction succeeded 0.2% of the time against GPT-6.2 Astra with updated safeguards, versus 54% against GPT-6 Astra with its original safeguards. That measures reasoning extraction, not every way to train a rival model.
Combined views
24.3K
2 Sources, first seen ago
Comments