Encrypted Reasoning Vulnerability Mitigated, but Plaintext Still Reachable via Model

lbeurerkellner · x · 2026-08-14

The author responds that the specific vulnerability is mitigated, but 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 →