Anthropic locks Claude’s chain of thought: it is invalidated if even a single line of context is modified, specifically to block model distillation.

Beating AI News Flash: Anthropic adds "context locks" to Claude’s hidden reasoning. Fable 5.1’s encrypted reasoning generated via API must now be returned exactly as it was when created, alongside the system prompt, tools, and historical messages used to generate it. If any preceding content is altered, the API will either throw an error or discard the reasoning segment entirely. This change is designed to prevent model distillation. Earlier, researchers found that while Claude’s encrypted reasoning is unreadable to users, it can be fed back to Anthropic’s compatible models. Attackers could prompt Opus to produce high-quality reasoning, then insert the encrypted block into the less secure Haiku model, tricking it into outputting Opus’s full reasoning process. This amounts to not only copying the large model’s final answer, but also stealing its step-by-step reasoning from the "scratch pad". When Anthropic launched Fable 5 in June, it already rolled out a fix by binding encrypted reasoning to the model, blocking it from being sent to smaller models like Haiku for decryption. Fable 5.1 now adds "conversation binding": as long as the model remains unchanged, modifying the preceding prompt, tools, or chat history will invalidate old reasoning. This lock is not yet fully enforced. It will only be mandatory for new API accounts created after August 31 when calling Fable 5.1; existing accounts are temporarily unaffected. Claude.ai, Claude Code, Cowork, and other Claude models are excluded from this update. Anthropic states that future new models will enable these rules by default for all users.