Encrypted Reasoning Traces Leak Across Models
Researchers report that encrypted reasoning blocks can be replayed across sessions, users and sibling models, allowing a weaker model to expose a frontier model's hidden reasoning. This is a concrete architectural failure rather than a generic jailbreak and could require per-session binding and strict isolation of opaque model state. It remains on watch because the apparent multi-source coverage traces back to one study; provider patches, independent reproduction or a disclosed exploit would confirm the broader infrastructure consequence.