Anthropic adds "context lock" to Claude's hidden reasoning to further prevent model distillation

Odaily News, Anthropic has begun adding a "context lock" to Claude's encrypted hidden reasoning. Fable 5.1's API-generated encrypted reasoning must remain consistent with the system prompt, tools, and message history at the time of generation. Once the context is modified, the related reasoning content becomes invalid.

It is reported that this mechanism is primarily designed to prevent model distillation attacks. Previously, attackers could potentially use other compatible models to attempt to read Claude's encrypted reasoning. This update further binds the reasoning content to the specific conversation context, building on the previous "model binding" to close the path of extracting the reasoning process by modifying the context.