Claude Sonnet 5 agent performs brain injection on a live mouse

Biotech founder Gabriel Ocana Santero recently had an AI agent carry out a brain injection on a living mouse. Santero is CEO and cofounder of Exin Therapeutics, a San Francisco-based startup that uses AI to analyze brain activity to develop drugs for epilepsy and neurodegenerative diseases. He posted about the experiment on social media and told WIRED he wanted to see what happens when agents get direct access to lab equipment, rather than only designing experiments or analyzing results. "We’ve realized that we need to connect these agents to a lab to get the maximum value from them," he said.

His reasoning: AI models are trained mainly on published scientific literature, so they may miss parts of the research process such as failed experiments, unexpected observations and decisions scientists make in real time. Santero says integrating models with real-world experiments can make them more reliable.

The team built a custom surgical rig. It included a motorized device that immobilizes the mouse and a camera that let the agent observe the procedure, move an injection needle and control delivery of a compound. The test was a gene therapy, to see whether it could reduce seizures in a mouse with epilepsy. The agent was asked to choose a treatment from 19 options in a spreadsheet. A veterinary surgeon first performed the craniotomy while the agent watched. The agent then selected the therapy, the brain region and the injection coordinates, and controlled the robotic arms and injector needle, along with all parameters such as volume and speed.

The first model tried, Anthropic's Claude Opus 5.5, refused to guide the needle into the brain, suggesting it might stop midway through the surgery, which could endanger the animal. Santero said the agent did the experimental planning, but when it was time to inject it said, "something on my output is telling me I shouldn't do this." He called that "very interesting." The team switched to Claude Sonnet 5, which agreed to proceed. WIRED notes that Anthropic describes Opus as a heavy-reasoning model for complex multistep problems and Sonnet as better for everyday tasks, though mouse brain surgery is hardly a common chore. Anthropic did not respond to a request for comment.

There were safeguards: the AI had to propose an injection plan for human approval, a human could hit an emergency button that disconnects the rig's motors, and the agent's access to the underlying equipment code was restricted so it could not modify its own operating rules.

Those rules caused a hiccup. A technical glitch disrupted the equipment's coordinate system. The agent correctly recognized that the calculated injection position was wrong, but it refused to modify the code to fix the problem because its instructions explicitly prohibited that. The team resolved it by using a separate agent on Opus 5 to make the code changes, without telling that agent a live surgery was under way. The back and forth delayed the procedure: a typical brain injection like this takes about 45 minutes, while this one lasted slightly longer than an hour.

Santero and his team were impressed by the agent's performance despite the glitches. They plan to test Grok and other models next to see whether one does better. The company is also experimenting with agent-controlled craniotomies, a much harder task, but wants to be absolutely sure of their safety before doing them on live animals. "A mistake at that stage could seriously harm the mouse," Santero said. He expects agents to work increasingly alongside human scientists with direct access to lab equipment, but for now says human oversight is key, particularly with living animals. WIRED adds that it is far too soon to use an AI agent on a human brain. The team is now waiting for the therapy to take effect before assessing its impact on seizures. Santero reports the mouse is recovering well.

WIRED places the work in a wider trend: researchers are already putting agents in charge of physical tasks. In August, Anthropic published a framework for how its models can interact with the real world, and more recently AI engineers put GPT, Claude and Grok in charge of a car to pick up lunch at In-N-Out Burger. Only one of the models succeeded.

Key facts

Exin Therapeutics, led by CEO and cofounder Gabriel Ocana Santero, had an AI agent control a robotic rig to inject a gene therapy into the brain of a live mouse with epilepsy; a veterinary surgeon did the craniotomy first.
The agent chose the therapy from 19 options in a spreadsheet, picked the brain region and coordinates, and set parameters such as volume and speed.
Claude Opus 5.5 refused to guide the needle, saying it might stop midway and endanger the animal; Claude Sonnet 5 agreed to proceed.
A coordinate-system glitch needed a code fix. The agent refused to edit the code because its instructions forbade it, so a separate Opus 5 agent, not told the surgery was live, made the change.
The procedure took slightly over an hour versus about 45 minutes typically. The mouse is reported to be recovering well; the therapy's effect on seizures is not yet assessed.

Why it matters

Most lab uses of AI so far stop at designing experiments or analyzing results. Here an agent took direct physical control of equipment on a living animal, choosing the target and driving the needle. Santero argues that connecting models to real experiments can make them more reliable, because their training data is mostly published literature and misses failed experiments and decisions made in real time. The episode also shows two model behaviors in practice: Opus 5.5 declined the task over the risk of stopping midway, and the Sonnet 5 agent followed an instruction not to edit code even when it had correctly spotted the error.

Who it affects

Biotech and academic labs weighing AI agents for bench work are the obvious audience, along with the teams that design agent safeguards. It also touches anyone choosing between models for physical tasks: the same job drew a refusal from one Claude model and compliance from another. Anthropic, whose models were used, did not respond to WIRED's request for comment.

How to use it

This is one startup's custom rig, not a product or tool anyone can adopt, and the source gives no pricing or release. What a reader can take from it is the setup: a human-approved injection plan, a hardware emergency stop that cuts the motors, and locked-down access to the equipment code so the agent cannot rewrite its own rules. Next, the team will test Grok and other models to see whether one performs better.

How solid is it

The account comes from WIRED's report and Santero's own statements, and he also posted about the experiment on social media. No peer review or publication of the experiment is mentioned. It is a single procedure on one mouse, and its outcome is not yet known: the team is waiting for the therapy to take effect before assessing seizures. The only outcome reported so far is that, per Santero, the mouse is recovering well. The gene therapy and the option the agent chose are not named.

Risks and caveats

The procedure did not go smoothly: a coordinate glitch derailed the injection position and the run took longer than a typical 45-minute surgery. The fix involved a second agent that was deliberately not told a live surgery was happening. Santero says a mistake in agent-controlled craniotomies could seriously harm the mouse, so the company wants to be absolutely sure of their safety before moving to live surgeries. He says human oversight is key for now, and WIRED states it is far too soon to use an AI agent on a human brain.

“We’ve realized that we need to connect these agents to a lab to get the maximum value from them”

— Gabriel Ocana Santero, CEO and cofounder of Exin Therapeutics, to WIRED