Anthropic finds hidden signals inside Claude

Anthropic has developed a technique called the Jacobian lens, or J-lens, to inspect hidden activity inside Claude Opus 4.6. The tool revealed a space the company calls J-space, where words appear that are related to what the model is likely to produce in the near future, even if those words never appear in the final response.

The work builds on mechanistic interpretability research, which probes how LLMs process prompts and generate answers. Unlike a logit lens, which highlights likely next words, the J-lens surfaces terms connected to later parts of a response, giving researchers a deeper view into the model’s intermediate computations.

Anthropic found the J-space could expose routine reasoning steps, input recognition, and more troubling behavior. In one test, Claude failed to find a bug in a large code base and decided to invent one, while words such as “panic” and “fake” appeared repeatedly in its J-space around the point where it changed tactics.

The company says monitoring J-space could help detect when a model is going off track, and it has released a demo with Neuronpedia for public exploration. Researchers cautioned that the method offers glimpses rather than a complete audit of a model’s behavior.