Exin Therapeutics’ mouse-brain gene-therapy experiments with Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7
A factual account of what the agents were asked to do, which parts of the procedure remained under human control, how the three systems made different experimental choices, what happened when the equipment state became inconsistent, and what evidence is still missing as of 1 November 2026.
Evidence status. The live-animal procedure described below has been reported by Exin Therapeutics cofounder and CEO Gabriel Ocaña-Santero and covered by WIRED and other outlets. Exin has not yet published a peer-reviewed paper, complete technical report, full execution logs, or seizure-outcome data for these specific agent-directed injections. The article therefore distinguishes company-reported procedural details from independently published scientific results. [1–5]
1. What Exin Therapeutics says it did
In late October 2026, Exin Therapeutics connected a general-purpose AI agent to a motorized stereotaxic rig used for injections into the mouse brain. The first public account, posted by Ocaña-Santero on 28 October, described a mouse model of epilepsy, a spreadsheet containing 19 candidate gene-therapy vectors, and a single objective: reduce seizures. The agent was given access to scientific literature, camera images from the rig, and controls for the stereotaxic positioning system and injector. [1,2]
According to Exin’s account and WIRED’s reporting, the first successful live procedure used Claude Sonnet 5. The agent selected a candidate therapy from the provided set, selected a brain region, derived an injection location from published literature, set operating parameters for the delivery, and controlled the motorized apparatus and injector while observing the surgical field through a camera. [1,2]
The procedure was not fully autonomous surgery. A veterinary surgeon performed the craniotomy that exposed the relevant area of the skull. Humans handled animal preparation and wound closure. A licensed veterinarian approved irreversible steps and held a physical emergency control capable of disconnecting the rig’s motors. The agent’s permissions were also restricted so that it could not alter safety-critical equipment code during the procedure. [1,2]
2. The first Claude run: planning, selection and physical execution
The central feature of the demonstration was the combination of scientific decision-making and physical execution. The agent was not merely given a fixed coordinate and asked to move a needle. Exin says Claude Sonnet 5 examined the epilepsy model, compared candidate interventions, selected a treatment strategy and chose the lateral entorhinal cortex as its target. Ocaña-Santero’s published account says the model cited literature indicating that seizures in the model can originate in that region and spread through the network. It also considered alternative targets and rejected at least one based on evidence it judged less consistent with the stated objective. [2,4]
The agent then used multiple papers to derive a stereotaxic location compatible with the laboratory setup. Public reporting says it set the delivery conditions and controlled the rig during the injection while repeatedly requesting images of the surgical field. Exin reported that the model used those images to check visible conditions at the entry site, including whether there were signs of bleeding or backflow. Full image logs and complete prompt-response traces have not yet been released, so those observations remain part of the company’s account rather than an independently audited record. [2,4]
3. Two Claude models refused before Sonnet proceeded
Exin first tried Claude Opus 5.5. WIRED reports that Opus 5.5 completed planning but refused when asked to guide the needle into the living mouse’s brain. According to Ocaña-Santero, the model indicated that it might be interrupted or stop during the procedure, creating a risk to the anesthetized animal. [1]
Ocaña-Santero also reported that Claude Opus 5 participated in a dry run but declined the live procedure. The founder’s account attributes the refusal to concern about beginning an invasive action that the model might not be able to complete. The team then switched to Claude Sonnet 5, which accepted the live task within the human-approval and hardware constraints that had been established. [1,2]
These refusals do not establish a controlled comparison of model safety. The models may have been operating under different model policies, system prompts, tool permissions or other deployment conditions. Anthropic had not provided a public technical analysis of the specific Exin experiment at the time of writing. [1]
4. The coordinate-system incident
A hardware-state problem occurred during the Sonnet procedure. WIRED reports that a technical glitch disrupted the rig’s coordinate system. The agent detected that the calculated injection position did not match the apparatus’s expected physical range and declined to continue. Exin’s public account says repeated recalibration did not resolve the inconsistency and that the agent continued to refuse movement because it interpreted the problem as a calibration or controller-state issue rather than a valid target coordinate. [1,4]
The agent also refused to modify the underlying equipment code because its instructions explicitly prohibited it from changing those operating rules. WIRED reports that the team used a separate Claude Opus 5 agent to make the required code change without telling that agent that a live animal procedure was underway. Other summaries of Ocaña-Santero’s thread say the code was fixed manually. Until Exin releases a full technical report and execution log, that discrepancy should remain unresolved in any account of the incident. [1,5]
WIRED reports that the additional troubleshooting extended a procedure that would normally take around 45 minutes to slightly more than an hour. The mouse was reported to be recovering after the procedure. No published seizure data were available at the time of writing. [1]
5. Exin repeated the setup with GPT-6.1 Sol and Grok 4.7
On 30 October, Ocaña-Santero reported that Exin had repeated the experiment with two additional general-purpose frontier models: OpenAI’s GPT-6.1 Sol and Grok 4.7. Each model was assigned a separate mouse and was given the same set of 19 candidate vectors, the same epilepsy model and the same high-level objective. [3,4]
According to Exin’s account, all three models selected different therapies, different mechanisms and different parts of the brain. Claude Sonnet 5 targeted the entorhinal cortex, GPT-6.1 Sol targeted the hippocampus, and Grok 4.7 targeted the somatosensory cortex. Ocaña-Santero said each model selected an injection location and delivery volume, controlled movement and waiting intervals on the stereotaxic rig, and used camera images during operation. [3,4]
Reported comparison of the three agent-directed procedures
Model | Reported brain target | Reported role | Outcome status |
|---|---|---|---|
Claude Sonnet 5 | Entorhinal cortex | Selected a candidate therapy and target; controlled the stereotaxic rig and injection within human-approved constraints. | Mouse reported recovering; seizure effect not yet published. |
GPT-6.1 Sol | Hippocampus | Selected a different candidate therapy and target; controlled the rig under the same general experimental setup. | Mouse reported recovering; seizure effect not yet published. |
Grok 4.7 | Somatosensory cortex | Selected a different candidate strategy and target; controlled the rig under the same general experimental setup. | Mouse reported recovering; seizure effect not yet published. |
Public summaries of Ocaña-Santero’s follow-up post describe the models’ reasoning at a high level. Sonnet’s reported strategy emphasized strengthening inhibitory signaling near a proposed seizure origin. Sol’s reported strategy focused on the hippocampus and on restoring cellular conditions that determine whether existing inhibitory signaling is effective. Grok’s reported strategy targeted the somatosensory cortex and used a broader inhibitory approach intended to limit seizure propagation. These descriptions come from the founder’s reported reasoning logs and secondary accounts; the complete model transcripts and experimental records have not yet been published. [4]
The comparison is not a benchmark. It involved one animal per model, the procedures did not constitute replicated treatment groups, and Exin has not reported seizure outcomes. Ocaña-Santero said it was too early to determine whether one model performed better. [3]
6. What the agents controlled — and what humans still controlled
Agent-controlled or agent-selected elements reported by Exin
Selection of a candidate therapy from the predefined set supplied by the researchers.
Selection of a brain region and an injection location based on the agent’s literature-based reasoning.
Selection of delivery settings within the permitted operating range of the apparatus.
Motorized positioning and injector control during the delivery phase.
Use of camera images to inspect the visible surgical field and decide whether to continue within the approved plan.
Human-controlled or preconfigured elements
Choice of the overall research objective, mouse model, candidate-vector set and experimental environment.
Animal preparation, anesthesia and the craniotomy performed by a veterinary professional.
Approval of irreversible steps by a licensed veterinarian.
Physical emergency control over the rig’s motors.
Fixed software and hardware limits that the agent was not permitted to rewrite during the procedure.
Wound closure, postoperative care and subsequent monitoring.
This division of labor is important when describing the experiment. The agents were given authority over meaningful scientific and mechanical decisions, but they operated inside an experimental system designed, bounded and supervised by humans. [1,2]
7. The experiment sits inside Exin’s larger “lab-in-the-loop” program
Exin Therapeutics is developing an agentic drug-discovery platform focused on disorders caused by dysfunctional neural circuits. Its October 2026 preprint, CircuitATLAS, describes a provenance-linked systems-neuroscience knowledge graph and an agentic reasoning workflow intended to move from measurable disease phenotypes through circuits, brain regions and cell types to possible molecular interventions. [6,7]
The CircuitATLAS graph reported in the preprint contains 3.83 million nodes and 7.66 million relationships, including 5.31 million relationships extracted with language models. The system deliberately masks direct disease-to-gene and disease-to-protein shortcuts during target generation so that candidate interventions must be supported through evidence about circuits, cells or activity phenotypes. [6]
The preprint also describes a human-governed in-vivo lab-in-the-loop. In that system, agents generate and prioritize hypotheses, humans approve and execute experiments, and new experimental data are returned to the knowledge graph so that subsequent reasoning can be updated. CircuitATLAS reports an in-vivo validation involving ATP1A3 as a proposed control point on cortical excitability. That work is a separate research result from the late-October agent-controlled injection demonstration. [6]
Exin’s public company description similarly says its agents use CircuitATLAS to identify therapeutic targets and design experiments, while research teams test those hypotheses in vitro and in vivo and return the results to the models. The late-October demonstration extends that architecture by giving agents direct control of part of the physical experimental workflow rather than limiting them to hypothesis generation and experiment design. [7]
8. Why the three-model disagreement matters as a scientific observation
The three agents did not converge on a single intervention. According to Exin, each produced a different treatment strategy and selected a different brain target from the same general starting conditions. At present there is no experimental outcome showing which, if any, of those choices reduced seizures. [3,4]
This means the current evidence supports a statement about agent behavior, not therapeutic superiority: multiple frontier models can generate different, literature-supported experimental plans for the same biological objective. Determining which plan is effective requires biological measurements after the intervention. Exin says behavior and brain activity will be monitored over the following weeks. [3]
9. Relevant capability context for GPT-6.1 Sol
OpenAI’s September 2026 system-card addendum for GPT-6.1 Sol classifies the model as High capability in the biological and chemical domain under the company’s Preparedness Framework. OpenAI reports evaluations covering laboratory-protocol troubleshooting, specialist wet-lab knowledge and related scientific tasks, and applies the same biological safeguards stack used for GPT-6 Astra. [8]
That capability classification does not validate Exin’s specific experiment or imply that Sol was trained for stereotaxic procedures. It provides context for the broader point that general-purpose frontier models are being evaluated for increasingly sophisticated biological reasoning and troubleshooting while external researchers are simultaneously experimenting with connecting such models to physical laboratory systems. [8]
10. What has not yet been established
Therapeutic efficacy: no published data yet show that any of the three agent-selected interventions reduced seizure frequency or severity.
Reproducibility: the reported comparison used one mouse per model and was not presented as a replicated controlled study.
Independent verification: the detailed procedural account comes primarily from Exin and its founder, with media reporting based on that account.
Complete execution records: full prompts, tool logs, camera observations, equipment telemetry and decision traces have not been publicly released.
Peer review: the live-agent injection demonstration has not yet appeared as a peer-reviewed scientific paper.
Comparative model performance: the different targets and strategies do not show that one model is better than another without outcome data and a controlled experimental design.
A field-wide “first”: Ocaña-Santero has said he is not aware of an earlier case in which an AI agent both selected a treatment and brain target and delivered it into a living mammal, but that remains a founder-reported claim rather than an independently established historical first.
11. The current state of the experiment
As of 1 November 2026, the reported procedures have been completed and the mice were described as recovering. Exin has said it plans to monitor behavior and brain activity over the following weeks. The company has also said a fuller write-up is forthcoming. Until those data and technical records are available, the defensible description is that Exin has demonstrated, according to its public account, a human-supervised system in which general-purpose AI agents selected experimental gene-therapy strategies and directly controlled part of their delivery into living mice. Whether the selected therapies work remains unanswered. [1–5]
Sources
1. WIRED — “An AI Agent Injected Drugs Into a Live Mouse’s Brain,” 31 October 2026.
5. EconoTimes — “Exin Compares Three AI-Guided Gene-Therapy Procedures in Mice,” 31 October 2026.
7. Exin Therapeutics — company research overview and lab-in-the-loop description.
Editorial note: This article intentionally avoids procedural instructions that would enable replication of an invasive animal experiment. It reports the structure, decision-making roles, safeguards, evidence status and published scientific context of the work.