A Startup Let an AI Agent Guide a Brain Injection in Mice — and Doesn't Know Yet If the Therapy Works
A startup company let an AI agent decide which gene therapy an epileptic mouse would receive, and where in the brain it would be injected — but kept humans in control of everything that could not be reversed. The entire account rests, for now, on the company's own thread on X, with no peer review and no data on whether the therapy works.
Epilepsy startup Exin Therapeutics of San Francisco connected a Claude agent to a motorized stereotaxic setup with cameras, gave it an epileptic mouse, and one goal: reduce the seizures. According to the company's CEO Gabriel Ocana-Santero, who described the episode in a thread on X on October 28, 2026, the agent chose the therapy and brain region itself, calculated the needle coordinates, and injected 250 nanoliters into the brain of the live animal. Wired and Tom's Hardware covered the experiment on October 31, and analysis firm XenoSpectrum followed up with a detailed breakdown of the division of roles on November 1.
The story has drawn attention for two reasons: first, a software agent made decisions that otherwise belong to trained scientists — choosing the therapy and target region in the brain. Second, Anthropic's most capable model, according to the company's own account, refused to do the job, while a smaller model went along. But what actually happened during the operation is worth separating carefully: what the agent did, and what the humans retained.
The Model That Said No
The team started with Claude Opus 5.5, which, according to Wired, would not steer the needle into the brain and warned that it could stop mid-operation and leave the animal in danger. "Starting something I can't finish on an anesthetized animal is worse than not starting," one of the Opus models is said to have told the team, according to Tom's Hardware — a statement known only as indirect paraphrase, not as a confirmed direct quotation. Claude Sonnet 5 then agreed to proceed.
Anthropic has not confirmed any of the model statements: the company did not respond to Wired's request for comment. And there is version uncertainty in the public sources — Martin Cid Magazine gives Opus 5.5 and Sonnet 5, while MSN writes only "Anthropic Sonnet." The details of the refusal should therefore be read as the account of the company and the reporting media.
What the Agent Decided — and What the Humans Kept
The account of the procedure itself is fairly precise about the division of responsibility. The agent was given a spreadsheet with 19 gene therapy vectors and three research papers, plus control of the equipment. It chose the vector and brain region — according to the XenoSpectrum analysis, the lateral entorhinal cortex, an area linked to epilepsy — and derived the coordinates from the papers. It also set the injection rate and waiting period.
What the agent was not allowed to do was everything irreversible. A veterinarian performed the craniotomy — the opening of the skull — while the agent watched; that part is, for now, off-limits to the agent. The veterinarian approved each irreversible step, and the agent had to present the injection plan for sign-off. A physical emergency stop could cut power to all the motors on the setup, and the agent was locked out of the equipment's code so it could not rewrite its own rules. The suturing was also performed by human surgeons.
The needle steering worked like this: the agent drove the motorized arms — the frame that holds the needle stable in three dimensions — and injected 250 nanoliters while monitoring the incision through the camera for bleeding. That said: this is the company's own description, conveyed through Ocana-Santero's X thread and secondary coverage.
When the Agent Spotted the Error — and Refused to Fix It
Along the way, an error occurred that disrupted the setup's coordinate system. The agent itself noticed that the calculated position was wrong, but refused to edit the code, because its instructions forbade it. The team brought in a second agent, powered by the older Claude Opus 5, to fix the problem — and did not tell it that an operation was underway in real time. The result was that a procedure that normally takes around 45 minutes stretched to a little over an hour.
The episode illustrates both the strength and the limitation of the setup: the agent caught the error, but the safety rules forced the humans to take over the correction — using a tool that did not know the consequences of its own task.
What We Actually Know — and Don't Know
The evidence base is thin, and any coverage of this story should state that explicitly. This is one mouse, described by the company itself on social media. XenoSpectrum concludes that all publicly available information rests primarily on the researcher's own account, and that no peer-reviewed article or technical report describing the experimental procedure has been found — nor any data showing therapeutic effect.
The efficacy question is not merely a formal caveat: the team is still waiting for the therapy to take effect before the seizures will be measured. No one knows, in other words, whether the treatment the agent chose actually worked. What can be said is that the operation appears to have gone well in line with the company's claim: according to information reported by MSN, the mouse has recovered well from the procedure and has been recorded moving around its enclosure.
Ocana-Santero has promised a fuller write-up on Substack. Until that appears, that is the basis available — the company's own account, amplified by several media outlets' reporting.
What Comes Next, and the Open Questions
Exin says it will test xAI's Grok and other models to compare them, and the company is working on agent-guided craniotomy — but wants to be sure it is safe before trying it on a live animal. That means the themes of this experiment will likely be repeated with more models and more autonomy.
At the same time, the company itself draws a clear line: letting an agent near a human brain, Ocana-Santero told Wired, is far too early. And the setup from the mouse experiment shows exactly why the boundaries are there — and why they were set: who is responsible when a software agent makes a mistake, or when a model refuses mid-procedure, remains an open question. So do the ethical frameworks for animal experiments in which the decisions are made by software.
The experiment is interesting not primarily because an agent steered a needle. It is that the entire design — locked code, signed plans, a physical emergency stop, an agent that refused to edit its own rules, and another that was allowed to do so without knowing what was at stake — represents a concrete attempt to answer what autonomy should actually mean in medicine. The answers, however, are not in yet.