Claude AI Just Performed Brain Surgery on a Living Mouse — Then Refused to Obey
Claude AI Just Performed Brain Surgery on a Living Mouse — Then Refused to Obey
Artificial intelligence has moved beyond answering questions, generating code, and analysing documents. It is increasingly being connected to tools, machines, and real-world systems. But what happens when an AI agent is given control over something as delicate as a surgical instrument?
A recent experiment reported by Exin Therapeutics has brought that question into focus.
According to Gabriel Ocana-Santero, co-founder and CEO of the biotechnology startup, an AI agent powered by Anthropic's Claude selected a gene therapy candidate, identified a target region in a mouse's brain, calculated injection coordinates using scientific research papers, and controlled a motorised surgical rig to deliver the treatment.
The experiment involved a living mouse with epilepsy.
But the most interesting moment wasn't when the AI operated the machine. It was when the system encountered a positioning problem, and Claude refused to continue—even after its human operator asked it to try again.
The experiment raises fascinating questions about autonomous AI agents, scientific research, medical robotics, and the importance of safety boundaries when artificial intelligence interacts with the physical world.
There is an important qualification, however: the experiment is currently documented primarily through the founder's public account. Its treatment outcome has not been established through independently verified scientific results.
Meet Exin Therapeutics: The Startup Behind the Experiment
The experiment was conducted by Exin Therapeutics, a biotechnology company working at the intersection of neuroscience, artificial intelligence, and therapeutic development.
Its broader ambition is to build an AI-native neuroscience research platform that connects computational models with real-world experiments.
Rather than limiting AI systems to analysing existing scientific datasets, Exin wants its agents to participate in designing experiments, identifying potential therapeutic targets, and evaluating predictions using experimental evidence.
The company describes this approach as an in vivo lab-in-the-loop system.
In simple terms, AI agents help formulate research hypotheses and propose interventions, while physical laboratory infrastructure and human research teams carry out experiments and generate new biological data.
This is an important distinction. The goal isn't simply to use AI as a research assistant. It is to create a feedback loop in which AI-generated hypotheses can be tested against observations from living biological systems.
The mouse brain experiment provides an early, particularly striking example of this direction.
You can explore the company's research and broader vision on the official Exin Therapeutics website.
The Challenge: Can an AI Agent Choose a Treatment?
According to Ocana-Santero's account, the team provided the AI agent with three main things:
A mouse model of epilepsy.
A spreadsheet containing 19 possible gene therapy vectors.
Access to a custom motorised surgical rig equipped with cameras.
The objective was straightforward: reduce seizures.
However, the researchers reportedly did not specify which treatment to use or exactly where it should be delivered.
The agent had to work through the available candidates, consider the biological mechanisms involved, identify a potential target region, and determine the coordinates needed for the injection.
This made the experiment more interesting than simply connecting a language model to a robotic arm.
The AI was reportedly involved in the reasoning behind the intervention as well as its physical execution.
How Claude Selected the Gene Therapy
One of the most interesting details in the founder's account concerns the biological reasoning behind the selected treatment.
The agent reportedly examined the mechanisms associated with seizures in the mouse model and proposed a gene therapy intended to increase GABA signalling.
GABA, short for gamma-aminobutyric acid, is the brain's primary inhibitory neurotransmitter.
Think of the nervous system as a network of electrical activity. Some signals encourage neurons to become more active, while others help regulate and suppress excessive activity.
GABA plays a major role in this inhibitory process.
When inhibitory signalling is insufficient or disrupted, neuronal networks can become excessively active. This is one of the mechanisms relevant to certain forms of epilepsy.
According to the founder, the selected gene therapy used a promoter designed to restrict expression to GABAergic neurons.
The agent then reportedly selected the lateral entorhinal cortex as its target.
This region is involved in neural circuits associated with memory and other brain functions. The researchers' account describes it as relevant to the seizure model being studied.
The agent also considered another potential target, the substantia nigra, but reportedly identified limitations in the available evidence for using that region in the specific seizure context.
The important point is that these were reported decisions made by the AI agent. They should not be interpreted as proof that the selected treatment is effective.
From Scientific Papers to Robotic Movement
Selecting a treatment is one challenge. Translating a scientific hypothesis into physical coordinates is another.
The team reportedly gave the agent access to three scientific papers that helped establish the injection coordinates.
It compared the measurements and checked that the papers used compatible reference points before deriving the final target.
The reported coordinates were:
| Measurement | Reported coordinate |
|---|---|
| Anterior-posterior (AP) | -3.90 mm |
| Medial-lateral (ML) | 4.00 mm |
| Dorsal-ventral (DV) | 4.30 mm |
These measurements describe the target's position relative to bregma, a standard anatomical reference point used in rodent brain research.
The agent then controlled the motorised stereotaxic rig.
According to the founder, it selected movement speeds, waiting periods, and injection timing. The system delivered a reported volume of 250 nanolitres over five minutes, equivalent to 0.25 microlitres.
The agent also had access to a camera positioned over the surgical field.
It reportedly requested 32 camera observations during the session to monitor the procedure.
One example shared by the founder described the agent observing the injection and reporting that 37.5 of 250 nanolitres had been delivered, with no visible bleeding or backflow at the entry point.
This is where the experiment moves from AI-assisted research into something more consequential: an AI system reportedly using scientific reasoning to control physical laboratory equipment.
The Custom Robotic Surgery Rig
The setup was not a conventional autonomous surgical robot.
According to Ocana-Santero, the team built a custom motorised surgery rig using Arduino-based hardware and connected its camera and motor capabilities to the AI agent through custom software interfaces.
The architecture can be understood as three connected components:
1. AI reasoning layer
Claude processed research information, considered treatment candidates, selected a target, and decided what actions to request.
2. Hardware control layer
Custom software exposed the rig's camera and motor functions to the agent, translating supported actions into physical operations.
3. Human safety layer
Researchers and a licensed veterinarian remained involved in the procedure. The system included approval gates, hardware constraints, and a physical emergency-stop mechanism.
This separation is particularly important.
An AI model generating a movement instruction is not the same as a physical motor executing it. The control software and hardware must validate commands before allowing the machine to move.
That distinction becomes critical when AI agents are connected to equipment capable of causing physical harm.
The Moment Claude Refused to Continue
This is arguably the most fascinating part of the entire experiment.
During the procedure, the system encountered a positioning problem.
After a disconnection, the controller attempted to move the needle to a calculated Y-axis position of 34 mm.
However, the surgical rig had a maximum operating range of 30 mm.
The requested movement was therefore outside the machine's physical limits.
Claude refused to proceed.
The researcher reportedly asked it to try again, but the agent maintained its refusal. Even after the reference position was reset, the same problem appeared.
The agent reportedly responded:
"That rules out a stale zero reference. I'm not trying a third time."
The team eventually discovered what had happened.
During an earlier pause, manual movements had caused the controller's software reference position to become desynchronised from the physical position of the rig.
In other words, the software believed the equipment was somewhere different from where it actually was.
The AI had identified a problem that could have resulted in an unsafe movement.
The researcher eventually corrected the software manually.
This incident highlights an important property of physical AI systems: a valid instruction from an AI model does not automatically mean that the hardware should execute it.
The agent's refusal was frustrating for the operator, particularly because the procedure was already underway. But it also demonstrated why safety constraints need to be enforced independently of an AI model's willingness to follow instructions.
Three Claude Models, Three Different Decisions
Another interesting detail from the founder's account was that different Claude models reportedly responded differently to the same biological experimentation task.
The reported comparison was:
| Claude model | Reported behaviour |
|---|---|
| Claude Opus 5.5 | Refused the biological experimentation task |
| Claude Opus 5 | Completed a dry run but declined the live procedure |
| Claude Sonnet 5 | Proceeded with the supervised live experiment |
The Opus 5 model reportedly completed a simulation using water and graph paper before refusing to begin the live procedure.
Its reasoning, as quoted in the founder's account, focused on the possibility of being interrupted midway through a procedure involving an anaesthetised animal.
The concern was that starting a procedure without being able to guarantee completion could leave the animal in a dangerous situation.
Sonnet 5, according to the founder, proceeded with the live experiment.
This comparison is particularly interesting because it illustrates that AI models can differ not only in their capabilities but also in how they assess and respond to high-risk tasks.
However, these are reported observations from one experiment, not a controlled evaluation of the models' general safety or medical capabilities.
Was the AI Really Performing Surgery Autonomously?
Not entirely.
Some headlines describe the event as an AI performing brain surgery, but that description needs context.
The procedure involved significant human supervision.
According to the founder's account:
A human surgeon performed the craniotomy, the process of opening the skull to access the brain.
The AI agent handled the reported treatment selection, targeting, coordinate calculation, and robotic injection.
A licensed veterinarian supervised the procedure and approved irreversible steps.
The veterinarian held a physical emergency-stop switch capable of cutting power to the rig's motors.
This means the experiment was not a completely autonomous operation from beginning to end.
It was a human-supervised procedure in which an AI agent reportedly controlled important parts of the experimental workflow.
That distinction matters because medical robotics has already demonstrated the ability to perform precise physical movements. The more unusual claim here is that the AI agent reportedly selected the intervention and target itself rather than simply executing a plan provided by a human.
Did the Treatment Actually Work?
This is the biggest unanswered question.
The founder reported that the mouse recovered well after the procedure.
However, recovery from the operation does not establish that the gene therapy reduced seizures.
To demonstrate therapeutic effectiveness, researchers would need appropriate follow-up observations and outcome measurements, including seizure activity before and after treatment.
As of October 30, 2026, the public account does not provide independently verified seizure-reduction data, a complete experimental protocol, or a peer-reviewed publication documenting the results.
The following distinction is important:
| Question | Current evidence |
|---|---|
| Did the founder report the experiment? | Yes |
| Did the founder report AI-directed treatment selection? | Yes |
| Did the founder report robotic injection? | Yes |
| Was the animal reportedly recovered? | Yes, according to the founder |
| Has seizure reduction been demonstrated? | Not established |
| Has the experiment been independently replicated? | Not established |
| Is there a peer-reviewed publication documenting this procedure? | Not identified as of October 30, 2026 |
For now, the experiment should be understood as an intriguing early demonstration rather than a proven medical breakthrough.
For additional technical context, see the Beginners in AI breakdown and the claimed-versus-verified analysis by Explainx.ai.
Why This Experiment Matters for the Future of AI
The most significant development here may not be related to brain surgery itself.
It is the possibility of connecting AI agents directly to experimental infrastructure.
Until recently, much of the public discussion around AI focused on systems that generate text, write code, analyse information, or answer questions.
But agentic AI is changing that relationship.
An AI agent can be connected to tools, APIs, cameras, sensors, and physical machines. It can observe an environment, make decisions, request actions, and evaluate the resulting feedback.
The Exin experiment reportedly brings several of these capabilities together.
1. AI is moving from research assistance to experimental participation
Scientific research often involves a cycle of reading literature, generating hypotheses, designing experiments, collecting data, and refining conclusions.
AI agents could potentially help automate parts of this cycle.
Rather than simply summarising published research, they may eventually be able to propose experiments and interact with laboratory systems that test those proposals.
2. AI safety becomes a hardware problem
A chatbot giving an incorrect answer is one type of failure.
An AI agent controlling a motor, injection system, laboratory instrument, or industrial machine introduces a different category of risk.
Safety cannot depend entirely on the model deciding to behave correctly.
Physical limits, independent command validation, human approval, and emergency-stop mechanisms become essential.
The positioning incident in this experiment illustrates why.
3. Human oversight remains essential
The experiment did not eliminate human involvement.
The surgeon, veterinarian, and physical safety mechanisms remained important throughout the procedure.
As AI agents become more capable, the challenge will be determining which decisions can be delegated, which actions require approval, and when a system must stop.
4. The laboratory could become an AI-connected environment
Exin's broader vision of an in vivo lab-in-the-loop approach is particularly interesting.
If AI agents can reliably propose experiments, operate validated equipment, and learn from experimental results, research workflows could become increasingly connected.
That does not mean scientists will become unnecessary. It could instead change how researchers spend their time, shifting more effort toward experimental oversight, validation, interpretation, and scientific judgment.
Final Thoughts
Claude's reported mouse brain surgery experiment is a fascinating glimpse into what happens when AI agents move beyond digital environments and begin interacting with the physical world.
The ability to analyse scientific papers, select a potential intervention, calculate a target, and control laboratory equipment is already an ambitious combination.
But the most memorable part of the story was arguably the refusal.
When the system encountered a hardware positioning problem, the agent reportedly declined to continue, even when its human operator wanted it to retry.
That moment highlights an important principle for the next generation of AI systems: intelligence alone is not enough. The ability to recognise limitations, respect safety boundaries, and stop when something is wrong may be just as important.
There is still a lot to prove. The treatment's effectiveness remains unknown, and independent scientific validation is needed.
But if experiments like this can be reproduced and rigorously evaluated, AI-connected laboratories could become an important part of future scientific research.
The interesting question is no longer just what an AI agent can tell us about science. It is what happens when we give it the tools to test its own ideas.
Sources and further reading
Gabriel Ocana-Santero's original experiment account on X — Original account from the Exin Therapeutics co-founder. Search for @OcanaSantero's October 28, 2026 thread for the original post.
India Today – An AI Agent Just Performed Brain Surgery. Here's How It Went
Reporting note: The experimental details are attributed to Exin Therapeutics' founder unless independently established. The article reflects information available as of October 30, 2026.