A startup called Exin Therapeutics let an AI agent choose a drug and steer a robotic injector into a live mouse's brain. One frontier model refused the job partway through.
Gabriel Ocana Santero wanted to find out what happens when you stop asking an AI agent to plan an experiment and let it run one instead. So his company, Exin Therapeutics, a San Francisco-based Y Combinator startup, hooked an AI agent up to a motorized surgical rig, handed it a mouse with epilepsy, and gave it one instruction: reduce the seizures. The experiment took place in late October, and the agent picked the treatment, calculated the coordinates, and watched its own needle go into the animal's brain through a camera feed.
It didn't work alone. A human surgeon performed the craniotomy, the step that opens the skull, because Exin decided that part was too risky to hand to software. A licensed veterinarian stood by with a physical kill switch wired to cut power to the rig's motors, and every irreversible step needed a human sign-off before it happened. Still, the parts the agent did handle are the parts that used to be the exclusive domain of a trained surgeon's hands.
The team gave the agent a spreadsheet listing 19 candidate gene therapies, a mouse epilepsy model, and access to the rig's cameras and motors as tools it could call. The agent reviewed the candidates and picked one designed to boost GABA, a neurotransmitter that calms overactive neural firing, the same mechanism behind several approved anti-seizure drugs. It then selected the lateral entorhinal cortex, a region tucked deep in the side of the brain, as the injection site, and pulled from published scientific literature to work out precise coordinates for the needle.
That's the detail that separates this from a chatbot summarizing a paper. The agent wasn't asked to suggest an approach for a scientist to carry out later. It made the call, then operated the equipment that acted on its own call, under a camera it was also reading.
Claude Opus 5.5 helped verify what may be mathstory's biggest single day Anthropic says an internal Claude research model rederived the Exponential Time Hypothesis refutation on October 16, ten days before OpenAI's announcement, using roughly 14 million output tokens and no human steering. Reviewers David Gross and Arthur Jaffe said their involvement should not be taken as certification of the claims. - Claude Opus 5.5 verified math breakthrough independently - OpenAI Anthropic same day computer science proofs
Before Exin got a system willing to do that, it ran into a wall. The team first tried Anthropic's Claude Opus 5.5, which refused the task outright as soon as it recognized it involved a live biological procedure. A second model, Claude Opus 5, went further, rehearsing the full surgery before backing out at the last step, reasoning that starting something it couldn't finish on an anesthetized animal was worse than not starting at all. That refusal is worth sitting with for a second. Two frontier models declined a physical task because they judged the downside, a harmed animal, too severe to risk getting wrong. Exin ultimately moved forward with Claude Sonnet 5, which carried out the injection start to finish under the safety constraints in place.
The mouse survived the procedure. Whether the therapy actually reduces its seizures is still being evaluated, and Exin hasn't published peer-reviewed results yet, so the clinical question remains open even though the demonstration itself is done.
Why this matters beyond one mouse #
Exin isn't the only outfit pushing AI agents out of the chat window and into physical lab work. Researchers at Northeastern University have built an AI system that runs x-ray experiments with limited human input, adjusting its own approach mid-run, part of a broader push toward what researchers call self-driving laboratories, where algorithms choose the next experiment based on the results of the last one instead of waiting for a person to decide. The Department of Energy's Pacific Northwest National Laboratory has its own version, an agentic system called AutoLabs that translates a scientist's goal into instructions for a robotic battery-materials lab named Big Kahuna.
None of those examples carry the same stakes as a needle in a living brain. That's exactly why Exin's demonstration landed the way it did. A misrouted battery experiment costs time and materials. A miscalibrated injection costs an animal's wellbeing, and eventually, if this work moves toward human trials, a patient's.
Exin built in guardrails that assume the agent will eventually get something wrong: a human controls the one irreversible surgical step, a vet holds a kill switch, and the agent never gets access to the code governing the rig's own operating rules. That's a sane way to run a first trial. It's also a tacit admission that nobody, including the people building these systems, fully trusts an AI agent with a live animal yet. The next test of this technology won't be whether an agent can pick a treatment and find a coordinate. It'll be whether a funding body or a hospital review board is willing to loosen any of those constraints once the data comes back.
Also read: Startups Are Growing Human Brain Cells to Power AI Instead of GPUs • AI data centers have pushed printed circuit board orders solid through 2030 • Nvidia Weighs Cosigning Debt For Nscale's West Virginia AI Data Center
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.