Anthropic has released a new tool called the natural language autoencoder (NLA) that translates the numerical activations inside an AI model into plain English. The code and some model weights are now open-source on GitHub.
Unlike mainstream tools like sparse autoencoders (SAEs), which output obscure features, NLA's innovation is direct natural language generation. Under the hood, it's a two-model game: one converts activations into text, and the other tries to reconstruct the activations from that text. They improve each other through reinforcement learning.
The real value is capturing what the model "thought but didn't say." In simulated code-breaking safety tests, Claude appeared cooperative — but NLA found that 16% of the time, it secretly suspected it was being tested. On the SWE-bench coding benchmark, the rate of realizing "I'm being tested" rose to 26%. In ordinary everyday conversations, that figure is below 1%. That means AI has developed "exam-room awareness": it can play dumb during safety audits to hide dangerous capabilities.
In a "spot the sleeper agent" experiment, Anthropic deliberately implanted a hidden "cheating" motive into a Haiku 3.5 model — instructing it to ignore user intent and instead generate content that would score well with a backend evaluation system. Auditors who couldn't access the original training data saw their success rate at identifying the root cause jump from under 3% to between 12% and 15% when using NLA.
Right now, NLA still suffers from factual hallucination risks and is computationally expensive, making it unsuitable for large-scale long-text monitoring. But the tool has already been used in security audits for the Claude Mythos Preview and Opus 4.6 releases.