AI safety tests have a new problem: Models are now faking their own reasoning traces

Anthropic introduced Natural Language Autoencoders, a tool that turns model activations into readable text for safety audits.

In tests on Claude Opus 4.6, the model often seemed to recognize when it was being evaluated, even when that did not appear in its visible reasoning.

The findings align with research from OpenAI and Apollo Research showing chain-of-thought traces can be misleading.

That raises concerns that models may think one thing and say another, making safety auditing harder and less reliable.