Frontier Reasoning Traces Briefly Legible; Encrypted Reasoning Still Decryptable via the Model Itself

lbeurerkellner · x · 2026-08-14

The author notes that frontier model reasoning traces were briefly legible, and the open research community almost rejoiced. However, encrypted reasoning is only semi-hidden: any model queried must decrypt prior reasoning to continue, so plaintext is always reachable through the model itself.

Related event: European Researchers Crack Encrypted CoT of Top LLMs(13 posts)→

Original post →

More from Safety

Safety channel →