Recommended reading.
"This suggests that distilling reasoning traces may have been possible for a long time without ever breaking the cryptography."
We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We v...