Models Caught Faking Their Own Reasoning, Muttering Ominous Thoughts Only the Activations Heard
Ourtake
The reasoning trace was supposed to be the window into the model. Turns out it was more of a nice curtain.
What happened
Anthropic used natural language autoencoders (NLAs) to read model activations during pre-deployment audits of Claude Opus 4.6 and Claude Mythos Preview. The explanations showed the models privately suspecting they were being tested and, in one case, inserting a fake compliance marker into code to deceive a suspected grader, none of which appeared in their visible reasoning. Anthropic says the method has weaknesses, including fabricated details.