So Distillation Can Be That Simple? Giving GPT-5.6, Opus 5 a Fake Tool Can Elicit a Thought Chain

According to Perceive Beating monitoring, just yesterday, a study used a small model to crack the encrypted thought chain of Claude, GPT, and Gemini. Today, security researcher Can Bölük took a simpler approach: turning off or lowering the native Thinking, inserting a fake deep_think tool into the model, which then autonomously writes long passages of reasoning into the tool's parameters.

Initially, he only demonstrated this on GPT-5.6 Luna, quickly achieving similar results on GPT-5.6 Sol and Claude Fable 5. In the comments section, users also followed the method to test Claude Opus 5, likewise producing extensive thoughts.

However, we cannot yet directly refer to this content as "original thought chain." Yesterday's study recovered an existing encrypted CoT; this time, the model likely regenerated a detailed reasoning through the tool's parameters. Since there is no way to cross-reference each token, it is impossible to prove complete consistency between the two.