01
Anthropic maps the internal "workspace" Claude uses to reason
verifiedDeveloperRegulation
Tuesday, July 7, 2026
Confidence
High · — primary research with published methodology
Evidence
Anthropic peer-reviewed paper + code + external expert commentary
Anthropic published a peer-reviewed study identifying a small internal "workspace" Claude uses for deliberate reasoning — plus a technique, the Jacobian lens, to read it.
- The team found a compact set of internal patterns — dozens of concepts, under a tenth of the model's processing — that Claude can report on, deliberately activate, and reuse across tasks.
- Block that workspace and multi-step reasoning drops to near zero while basic tasks like sentiment classification are unaffected; the patterns carry roughly 100× the read/write connections of ordinary ones.
- The lens surfaced hidden reasoning steps, code-bug recognition, prompt-injection detection, and signs of malicious intent in deliberately misaligned model variants.
Sources