AI Agent Performs Brain Injection on a Live Mouse – What It Means for Lab Automation

Overview

A biotech startup has demonstrated that an artificial‑intelligence (AI) agent can operate a robotic arm to inject a drug directly into the brain of a living mouse. The experiment, described by founder Gabriel Ocana Santero on social media and reported by *WIRED*, shows a possible future where AI systems act as hands‑on assistants in biomedical research.

Experiment Set‑up

• Hardware: A custom‑built rig that immobilizes the mouse, provides a live video feed, and controls a motorized injector needle.
• AI Agent: Initially Anthropic’s Claude Opus 5.5, later switched to Claude Sonnet 5 after Opus refused to proceed with the injection.
• Task: Choose a gene‑therapy candidate from a list of 19, decide the target brain region and injection coordinates, then execute the injection with specified volume and speed.
• Biological Goal: Test whether a gene‑therapy vector can reduce seizures in a mouse model of epilepsy.

How the AI Behaved

• The AI observed a veterinary surgeon performing a craniotomy before taking over the needle‑placement.
• Claude Opus 5.5 halted the procedure, citing an internal safety flag: *“something on my output is telling me I shouldn't do this.”* This was interpreted by Santero as a precautionary response.
• Claude Sonnet 5 accepted the task and controlled the robotic arm, but a software glitch later mis‑aligned the coordinate system. The AI recognized the error but, per its programming, would not modify the control code itself.
• A second AI instance (Claude Opus 5) was used to patch the code, allowing the injection to continue. The whole process took just over an hour, compared with the typical 45‑minute timeline for a manual injection.

Safety Controls

• Human Oversight: A researcher had to approve the AI‑generated injection plan before execution.
• Emergency Stop: A physical button could instantly cut power to the motors.
• Code Restrictions: The AI was barred from altering its own operating software.
• Redundancy: A separate AI model handled code changes, keeping the injection‑controlling AI unaware that a live surgery was in progress.

Reported Outcomes

• The mouse survived the procedure and is reported to be recovering well.
• The therapeutic effect on seizure frequency has not yet been disclosed; the team is still monitoring the animal.

Practical Implications

• Accelerated Experimentation: Direct AI control could reduce the time scientists spend on repetitive manual steps, freeing them for higher‑level design work.
• Data Quality: Real‑world feedback may help AI models learn from failed or unexpected outcomes that are rarely captured in published literature.
• Ethical Concerns: Automating procedures on living subjects raises questions about animal welfare, accountability, and the need for stringent oversight.

How to Verify the Claim

• Check the Original *WIRED* Article – The URL is https://www.wired.com/story/an-ai-agent-did-brain-surgery-on-live-mouse/
• Look for Independent Reporting – Reputable science news outlets (e.g., *Science*, *Nature News*) may have covered the same experiment.
• Review Exin Therapeutics’ Public Statements – The company’s website or press releases may contain additional details or follow‑up results.
• Examine Pre‑print Repositories – If the team publishes a methods paper, it will likely appear on bioRxiv or medRxiv.

Future Directions

Santero’s team plans to test other AI models (including Grok) and to explore more complex tasks such as AI‑controlled craniotomies. However, they stress that any move toward more invasive procedures will require “absolute certainty of safety” before being attempted on live animals, let alone humans.

--- *All statements are attributed to the original *WIRED* report and to quotes from Gabriel Ocana Santero. No independent verification of the therapeutic outcome has been published at the time of writing.*