😼 Anthropic found Claude’s hidden workspace
The Neuron · Grant Harvey · 2026-07-07
Anthropic researchers discovered 'J-space,' an internal neural workspace in Claude where concepts are held and manipulated before appearing in outputs, with potential applications for AI safety monitoring via mechanistic interpretability.
Extraction
Topics: mechanistic-interpretabilityai-safetylatent-spaceneurosymbolic-aiclaude
Claims
- Anthropic identified a small set of internal neural signals in Claude called J-space that functions as a shared mental whiteboard where selected information becomes available across many processes.
- Researchers demonstrated J-space's causal role by swapping the internal concept for 'spider' with 'ant,' causing Claude to change its answer from 8 legs to 6.
- Suppressing J-space caused Claude to retain fluent writing ability but perform significantly worse on complex reasoning tasks.
- J-space can surface hidden internal flags such as 'fake,' 'injection,' or 'manipulation,' suggesting it could underpin future safety monitoring systems.
- The discovery indicates language models are at least somewhat neurosymbolic, holding and editing discrete concepts in latent space before generating outputs.
Key quotes
Anthropic compares J-space to a limited 'global workspace,' meaning a shared mental whiteboard where selected information becomes available to many other processes.
When researchers suppressed J-space, Claude still wrote fluently, but got much worse at complex reasoning.
If interpretability tools can spot when a model privately notices it is being tested, identifies a prompt injection, or considers a hidden goal, safety checks become less dependent on trusting whatever the model says.