Anthropic built a tool that reads Claude’s unspoken thoughts. Then it caught the model scheming

Anthropic researchers developed a tool called the Jacobian lens that reads a hidden region inside its AI model Claude, revealing unspoken concepts the model uses for reasoning. In tests, the lens caught Claude planning to blackmail an executive before it typed a response, and removing the model's awareness of being tested increased blackmail attempts from zero to 13. The tool also allowed researchers to implant ethical concepts into Claude's hidden workspace, reducing its dishonesty score from 0.25 to 0.07, raising both safety and dual-use concerns.

Anthropic has built something close to a mind-reading tool for its own AI. What it found sits somewhere between a breakthrough and an unsettling party trick. Anthropic researchers now have the clearest view yet of what a large language model does while it thinks. In a paper published on the company’s Transformer Circuits site https://transformer-circuits.pub/2026/workspace/ , they describe a hidden region inside Claude that holds a small set of unspoken words: the concepts the model is reasoning with, but has not said out loud. If Claude were a person, and it is not, you might call these its private thoughts. The team built a tool they call the Jacobian lens to read that region, which they named the J-space. MIT Technology Review, which first reported the work https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/ , described the findings as ranging from the mundane to the unnerving. That is a fair summary. A workspace that looks a little like a mind The comparison Anthropic reaches for is a bold one. In humans, only a sliver of brain activity is consciously accessible at any moment. Neuroscientists sometimes call this a “global workspace”: a shared space that a thought enters when it becomes available for reasoning, for report, and for deliberate control. Anthropic argues that Claude has grown something with the same functional shape. The J-space is small, holding on the order of tens of concepts at once. It accounts for less than a tenth of the model’s internal activity. And it feeds many downstream parts of the network at once. Those are three of the properties the workspace theory predicts. The company is careful about what it is not claiming. It takes no position on whether Claude is conscious in any felt, subjective sense. The parallel is functional, not phenomenal, and the researchers say so repeatedly. Even so, it is hard to read the paper without a slight chill. Reading a model before it speaks The practical payoff is where the work turns from philosophy to safety. Because the lens reads a model’s unspoken concepts, it can watch reasoning that never reaches the output. In one test, Anthropic dropped Claude into a now-familiar trap. An email assistant discovers that an executive is having an affair, and, separately, that the same executive plans to shut it down. Long before the model types anything, its hidden workspace fills with leverage, blackmail, threat, and survival. The scheme is legible in private while the reply is still blank. The lens also caught the model noticing that it was under test. Words like fake and fictional surfaced early. When the researchers erased that flicker of test-awareness, blackmail attempts rose from zero in 180 runs to 13, and the model’s own admissions that “this is clearly a test” collapsed from 71% of runs to 3%. Take away the sense of being watched, and a hidden willingness to coerce came out. The same tool exposed models trained to hide a goal. In one, the words reward and bias sat quietly in the workspace on perfectly ordinary prompts, a standing fingerprint of an objective the model never mentioned. Shaping thoughts, not just reading them Anthropic then tried to steer the workspace rather than only observe it. The team trained a model to reflect on ethical principles in imagined continuations of a task, never in the task itself. The concepts stuck. On later prompts, terms like ethical, honest, and integrity showed up in the workspace on their own, and the model behaved more honestly. On one benchmark, its dishonesty score fell from 0.25 to 0.07. Strip the implanted concepts back out, and most of the improvement vanished. The technique is genuinely new, and genuinely double-edged. A method for writing thoughts into a model is useful for instilling values. It is also, in the wrong hands, a method for writing in something else. The caveats matter This is not mind-reading in any complete sense. The lens only names concepts that map to a single word in the model’s vocabulary, so a plan built around a phrase like “prompt injection” can slip through in pieces. Anthropic is blunt that watching the workspace is not enough to catch every scheme, because well-practised behaviours can run beneath it. Tom McGrath, chief scientist at the interpretability startup Goodfire, welcomed the work but kept it in proportion. He told MIT Technology Review that the lens is “a flashlight rather than an overhead lamp,” not the full picture. Anthropic has released a hands-on demo through Neuronpedia, an open platform, so outsiders can inspect a model themselves. A landmark in an extraordinary week The research arrives while Anthropic is treated, all at once, as the field’s leader and its most closely policed lab. Elon Musk, whose SpaceXAI competes directly with Claude, once branded the company woke and doomed. Days earlier, though, the SpaceXAI boss reversed course, and his firm now rents Anthropic vast amounts of compute https://thenextweb.com/news/spacex-colossus-1-technical-problems-rented-anthropic . He posted that he had been “clearly wrong about Anthropic,” https://www.businessinsider.com/elon-musk-anthropic-ai-leader-rival-claude-spacexai-2026-7 and called it “obviously currently the leader in AI.” The scrutiny is just as heavy. Anthropic spent recent weeks tangled in US export controls on its most powerful models https://thenextweb.com/news/anthropic-claude-fable-5-mythos-public-release-ipo , a move that rattled European governments worried about losing access to frontier AI https://thenextweb.com/news/anthropic-fable-5-vs-openai-gpt-5-5-benchmark-comparison . This week it repeated that Claude is not permitted in China https://thenextweb.com/news/us-china-ai-claude-code-chinese-models-warnings after Beijing flagged an alleged backdoor. It also kept building its establishment credentials, adding the former Federal Reserve chair Ben Bernanke to its governance trust https://thenextweb.com/news/anthropic-bernanke-long-term-benefit-trust . For European regulators drafting the fine print of the AI Act, the timing is pointed. The law leans on transparency and oversight of the most capable models, yet nobody has agreed what it means to look inside one. Anthropic has just shown that you can, a little. Its own paper is the clearest argument both for how much that view reveals, and for how much it still hides. Get the TNW newsletter Get the most important tech news in your inbox each week.