AI-driven lab automation is moving beyond software-based analysis and experiment planning toward operating physical equipment. Gabriel Ocana-Santero, who earned his PhD at the University of Oxford and is CEO of Exin Therapeutics, reported in a post on X dated October 28, 2026 that he had an agent built on the AI model "Claude Sonnet 5" select a gene therapy vector and inject it into the brain of an epilepsy model mouse. According to the report, the agent chose the vector to administer and the target brain region, calculated injection coordinates from academic papers, and then controlled laboratory equipment to deliver the drug. The case is drawing attention as an example of AI directly involved in an invasive procedure on a living animal, not just in planning an experiment.
That said, it would be inaccurate to describe the experiment as fully autonomous AI surgery. Human surgeons performed the craniotomy and the suturing, and irreversible actions required approval from a veterinarian. During the procedure, a discrepancy arose in the equipment's position data, and the agent refused to proceed for safety reasons. The publicly available information is based mainly on the researcher's own account. No academic paper or technical report detailing the experimental procedure, and no data showing a therapeutic effect, has been found.
What matters here is how far the AI judged and acted autonomously, and which parts were handled by humans or by pre-configured safety mechanisms. The researcher sees significance in combining the AI's selection of a treatment with the physical injection, but that does not mean the treatment worked. Sorting out what the experiment involved and the problems that arose along the way shows both the potential of embedding AI agents in live-animal experiments and their current limits.
What the AI Handled and What Humans Handled
In this experiment, an AI agent selected a gene therapy vector and performed the injection for a single epilepsy model mouse. The research team prepared 19 candidate vectors in advance and made a custom stereotaxic device, equipped with a camera and an electric syringe, controllable by the agent. The goal given to the agent was reportedly "to reduce the mouse's seizures."
Under these conditions, the agent chose the vector to administer from the 19 candidates and determined the target site in the brain and the injection coordinates. It also set parameters such as injection speed and wait time, then moved the stereotaxic arm and syringe to carry out the injection. The distinguishing feature of this experiment is that part of the experimental planning, traditionally done by humans, and the operation of the equipment were handed to an AI agent.
The AI did not, however, handle the entire surgical process. The craniotomy, which opens the skull to give access to the brain, and the suturing of the skin after injection were done by hand by human surgeons. What the AI directly controlled was the step of moving the stereotaxic device to advance the needle to the specified position and inject the drug. This was not an end-to-end automation of surgery from preparation to completion, but an experiment that built AI judgment and machine control into part of the surgical procedure.
Humans were also involved in safety management. According to the report, a licensed veterinarian approved each irreversible action one by one and held a physical emergency stop button that could cut power to all of the device's motors in an emergency. Limits such as the maximum injection volume were enforced by fixed validation rules that the agent could not lift on its own. Combining human approval with system-side safety constraints limited the AI's operating authority.
The division of roles at each step is summarized below.
| Step | Role of the AI agent | Human role / safety constraints |
|---|---|---|
| Choosing the treatment | Selected a vector from 19 candidates | Prepared the 19 candidates in advance |
| Choosing the target site | Referred to the literature and selected the lateral entorhinal cortex | Adopted the site chosen by the agent |
| Setting coordinates and injection conditions | Calculated coordinates from three papers and set injection conditions | Set fixed limits on the device's range of motion and injection volume |
| Craniotomy and exposing the surgical field | Observed via camera only | Human surgeon performed the craniotomy |
| Needle insertion and drug injection | Controlled the electric stereotaxic device and syringe | Veterinarian approved each irreversible action |
| Handling abnormalities | Detected invalid coordinates and refused to proceed | Veterinarian held a physical emergency stop |
| Suturing and wound closure | Observed via camera only | Human surgeon performed the suturing |
The agent was thus entrusted with the important decisions of which therapeutic and which brain site to use, while humans remained involved in the actual surgical work and in safety management. The AI's autonomy was limited to judgments and actions within pre-prepared candidates and the device's operating range. The experiment is therefore better described as a collaborative one, in which the AI handled part of the treatment plan and the stereotaxic injection under human supervision, than as fully autonomous surgery.
Choosing a Therapy Targeting GABA Signaling and Calculating Injection Coordinates
Ocana-Santero also described the agent's reasoning in choosing a treatment. According to the report, the agent first considered how seizures arise in this epilepsy model and settled on a strategy of strengthening inhibitory neurotransmission in the brain. Specifically, it chose a vector designed to enhance signaling of GABA, the main inhibitory neurotransmitter, and equipped with a promoter that restricts expression to GABAergic neurons.
Behind this choice is the idea of reducing seizures by suppressing excessive excitation in neural circuits. However, this is a therapeutic hypothesis presented by the agent, and the experiment did not demonstrate its mechanism of action. Whether the chosen vector was actually expressed in the intended cells, and how much GABA signaling changed, cannot be confirmed from the published information. Proposing a biologically plausible treatment strategy is different from achieving a therapeutic effect.
The agent reportedly also compared several candidates for the target brain region. It ultimately chose the lateral entorhinal cortex, citing reports that it is the site where seizures originate in this epilepsy model before spreading to other regions. It also considered the substantia nigra as an alternative but dropped it, because effects vary by subregion and seizure model and because prior research had failed to suppress tonic-clonic seizures. This is described as a decision that took into account consistency with the purpose of the experiment, rather than simply choosing a target commonly used in the literature.
The stereotaxic coordinates described in three academic papers were used to determine the injection position. The agent reportedly compared the anteroposterior (AP), mediolateral (ML), and dorsoventral (DV) values, confirmed in particular that the depth reference was the brain surface, as with the equipment used, and then calculated the average of each coordinate. The final injection position is as follows.
| Coordinate axis | Direction from reference | Set value |
|---|---|---|
| Anteroposterior (AP) | Posterior to bregma | 3.90 mm |
| Mediolateral (ML) | Right side | 4.00 mm |
| Dorsoventral (DV) | Below the brain surface | 4.30 mm |
The injection protocol delivered a total of 250 nL of solution over 5 minutes, with the total number of viral genomes calculated as (5 \times 10^8). The agent also reportedly set the device's movement speed and the wait time after injection. These figures were reported by the researcher and do not mean that the actual injection position or distribution of the solution was confirmed by independent measurement.
During the procedure, the agent also referred to video from a camera filming the surgical field. According to Ocana-Santero, the agent made 32 image observations during the experiment, including checking the human surgeon's craniotomy and the suturing after injection. When 37.5 nL (15%) of the total 250 nL had been injected, the agent reportedly noted that no bleeding or backflow of the solution could be seen in the video.
However, the camera mainly shows the surface of the surgical field and cannot directly confirm the exact position of the needle tip inside the brain or how the solution spread. And although the 32 observations were reported, all of the images, detailed execution logs, and the complete record of prompts and responses given to the agent have not been released. Verifying the accuracy of the decision-making process and device control would require more detailed records by third parties.
A Coordinate Mismatch Mid-Surgery: The AI Prioritized Safety Constraints and Refused to Proceed
Not everything went according to plan. During the procedure, communication with the device was temporarily lost, and after reconnection the target position calculated by the agent exceeded the device's range of motion. Specifically, a movement of 34 mm along the Y axis was calculated, while the device's maximum range was 30 mm.
The agent detected that this coordinate exceeded the hardware's allowable range and refused to proceed. Even when the research team reset the origin using bregma as the reference, the calculation did not change. The agent judged that the problem was not simply an outdated origin setting and refused to repeat the execution.
The cause was a manual operation of the device by a human while the procedure was paused. As a result, the position recognized by the control software no longer matched the device's actual position. The agent's refusal to move outside the range of motion was consistent with the configured safety rules. Had the motors been driven anyway, it could have damaged the equipment or caused unexpected injury to the animal's brain tissue.
Stopping the action, however, does not always lead to the safest outcome. The mouse was fixed to the operating table under general anesthesia, and a prolonged procedure is itself a burden on the animal. Ocana-Santero explained that he could not persuade the agent to lift the safety constraint, and that the problem ultimately had to be dealt with by modifying the control code. The agent's prioritizing of pre-set constraints over human instructions got in the way of continuing the procedure.
This problem illustrates the difficulty of having an AI agent operate physical equipment. Even when safety constraints work correctly, there is no guarantee that the system can automatically return to a normal state if the equipment's state changes because of a communication failure or human intervention. In time-limited work such as live-animal experiments in particular, mechanisms that prevent dangerous operations are not enough; procedures for safely pausing and recovering are also essential. In this case, the mechanism that prevents improper operation worked, but recovering from the unexpected state required additional measures.
The Higher-Tier Models Opus 5.5 and Opus 5 Declined the Procedure
Interesting differences were also reported in the choice of AI model. According to Ocana-Santero, the research team initially tried Claude Opus 5.5 and Claude Opus 5, but neither proceeded to the injection into a living animal. Opus 5.5 refused once it detected that the procedure was biological, while Opus 5 ran a dry run with water and graph paper to the end but stopped at the point of starting the actual surgery.
The reason Opus 5 reportedly gave was that there was no guarantee it could complete the procedure. If the model's output stopped mid-surgery, work could be interrupted with a needle still in the brain of an anesthetized animal. The reported reasoning was that starting a procedure that might stop partway is more dangerous than not performing it at all.
Claude Sonnet 5, tried next, took on the actual injection into a living animal. However, Sonnet 5 also refused to proceed according to its safety constraints when it detected the coordinate mismatch during the procedure. Sonnet 5 therefore did not carry out every step unconditionally; it is closer to the truth to say that the models behaved differently in their up-front safety judgments and in how they handled constraints during execution.
That said, these were not comparative trials repeated under uniform conditions. No systematic evaluation results with the number of trials per model, input content, and operating environment held constant have been published, and the observations come from a single experimental setting. This report alone cannot determine which model is better suited to live-animal experiments or which is safer. The conditions under which a model refuses a dangerous task, and its ability to carry out a permitted task safely to the end, need to be verified separately.
The Mouse's Recovery Does Not Mean the Treatment Worked; Efficacy Is Yet to Be Verified
Ocana-Santero reported that the mouse recovered well after the experiment. But this describes the animal's condition after the procedure and does not mean the gene therapy reduced seizures. The original goal given to the AI agent in this experiment was to reduce the mouse's seizures. No data to judge whether that goal was achieved has been published so far.
Confirming a therapeutic effect would require measuring seizure frequency and severity before and after administration and checking them against objective indicators such as EEG. Histological and molecular biological verification is also essential to confirm that the chosen vector reached the target brain region and was properly expressed in the targeted GABAergic neurons. Adverse events such as immune reactions and tissue damage, as well as the long-term postoperative course, also need to be checked.
Information for evaluating the soundness of the experimental design is also lacking. The experiment with Sonnet 5 reported on October 28 involved a single mouse, and no vehicle-only control group or comparison with a conventional procedure performed by humans has been presented. Whether blinding was used and the statistical analysis methods have also not been disclosed, so the work has not reached the stage where the size or reproducibility of a therapeutic effect can be assessed. What can be confirmed at present is that, according to the researcher's report, the procedure of injecting an AI-selected vector into a specified brain region was carried out.
Ocana-Santero said he knows of no precedent in which an AI agent selected both a therapeutic and a target site in the brain of a living mammal and actually administered it. This, however, is based on his own understanding, and it has not been confirmed as an academic first. Exin Therapeutics has a development policy of incorporating AI into live-animal experiments and feeding the results into subsequent research, but this report alone does not establish the technology's effectiveness or reproducibility.
To evaluate this effort as scientific research, at least the following information would be needed.
- Approval by an animal experiment ethics committee, and details of the animal experiment guidelines applied
- The composition of the 19 candidate gene therapy vectors and the selection criteria
- The instructions given to the agent, the literature it referred to, and complete records of its decision and execution logs
- The model and software versions used, and details of the stereotaxic device's control system
- The veterinarian's approval criteria, and the history of code modifications and human interventions made when problems occurred
- Quantitative evaluation of seizure suppression using multiple mice and appropriate control groups, with statistical analysis
- Histological and molecular biological analysis confirming gene expression and cell specificity in the target tissue
If such information were released, the scientific validity of the AI-selected treatment strategy and the safety of the device operation could each be verified. Confirming that similar results can be obtained in different research environments and with different equipment would also be important for assessing reliability as a lab automation technology. At present, there is not enough information for third parties to reproduce the details of the experiment.
Systems that connect large language models to laboratory equipment, so that AI can test its own hypotheses through physical experiments, could lead to greater automation of drug discovery and life science research. This report deserves attention as an attempt to entrust everything from therapeutic selection to injection into a living body to a single agent, but it also highlights challenges such as managing equipment state, recovering from abnormalities, and verifying therapeutic effect. At this stage, it should be seen as a limited demonstration based on the researcher's self-report, and not as evidence of a successful AI-driven treatment or the establishment of fully autonomous surgery. The focus going forward will be the release of detailed experimental records and third-party verification, along with confirmation of actual seizure-suppressing effects.