According to CoinMeta, Anthropic has started to add a "context lock" to the hidden thoughts of Claude. The encrypted thoughts generated by Fable 5.1 through API must be transmitted back in their original form, along with the system prompts, tools, and historical messages used during generation. If any part of this content is altered, API will report an error or simply discard that thought. This change is intended to prevent model distillation. Researchers have found that although users cannot understand the encrypted thoughts of Claude, they can still be read by compatible models such as Anthropic. Attackers can first have Opus produce high-quality reasoning, then pass on the encrypted blocks to a less protected Haiku, inducing it to reveal the complete thoughts of Opus. This is equivalent to not only copying the answers from large models but also taking away the problem-solving processes written on scratch paper. This time, with Fable 5.1, a "conversation binding" feature has been added; any modification to the previous prompts, tools, or chat records will render the old thoughts invalid as well.