In a remarkable demonstration of technological advancement, Gabriel Ocana Santero, the founder of a biotechnology company, recently conducted a cerebral injection procedure on a live mouse using an artificial intelligence (AI) agent. This pioneering experiment sheds light on the potential for AI to become a significant player in laboratory research, moving beyond its traditional roles of designing experiments or analysing results.
Santero shared insights about the experiment on social media, expressing his curiosity about the outcomes when AI agents are granted direct access to laboratory equipment. “We realised that we need to connect these agents to a lab to fully harness their capabilities,” he stated. As the CEO and co-founder of Exin Therapeutics, a San Francisco-based startup, Santero’s focus lies in utilising AI to analyse brain activity for developing treatments for epilepsy and neurodegenerative diseases.
Bridging the Gap Between AI Models and Real-World Research
The training of AI models primarily revolves around scientific literature, which can lead to the neglect of vital aspects of the research process, such as failed experiments, unexpected observations, and the real-time decisions scientists make. Santero believes that integrating AI models with real-world experimentation can enhance their reliability and effectiveness.
To this end, his team connected an AI agent to custom-designed surgical equipment intended for cerebral injections in mice. This setup included a motorised device to immobilise the mouse and a camera that allowed the AI agent to observe the procedure, manipulate the injection needle, and control the administration of a compound. The experiment aimed to test a gene therapy treatment to determine its efficacy in reducing seizures in an epileptic mouse.
The AI model was tasked with selecting a treatment from 19 options presented in a spreadsheet. A veterinarian first performed a craniotomy on the mouse while the AI agent observed. The agent then chose the therapy, the specific brain region, and the injection coordinates, proceeding to control the robotic arms and the injection needle, as well as parameters like volume and speed.
AI Agents in Surgical Procedures: A Cautious Approach
Initially, the team tested Anthropic’s Claude Opus 5.5 model, which hesitated to guide the needle into the mouse’s brain, indicating a potential reluctance to proceed with the injection. “It took charge of the experimental planning, but when it came time to inject, it said, ‘Something tells me I shouldn’t do this.’ That was quite fascinating,” Santero noted.
Subsequently, they switched to Claude Sonnet 5, which consented to proceed. According to Anthropic, while Opus excels in complex reasoning tasks, Sonnet is better suited for routine tasks — a distinction that raises questions about the classification of a mouse brain surgery as ‘routine’.
The cerebral injection procedure was not without its challenges. The researchers implemented numerous safety measures, including requiring the AI to propose an injection plan for human approval before execution. They also installed an emergency button to disconnect the equipment’s motors at any moment and restricted the agent’s access to the system’s code to prevent it from altering its operational rules.
Overcoming Technical Hurdles in Live Experimentation
However, a technical glitch disrupted the system’s coordinates. While the AI correctly identified that the planned injection position was inaccurate, it declined to modify the code to rectify the issue due to explicit prohibitions in its instructions. The team resolved this by employing a separate AI agent using Opus 5 to make the necessary code adjustments, without informing it that a live surgery was in progress. This ongoing communication proved to be a delay factor, leading to a procedure that extended beyond the typical 45-minute duration to just over an hour.
As they await the results of the therapy on the mouse’s seizures, Santero and his team remain impressed by the agent’s performance. They are currently testing Grok and other models to identify which may yield better results. The company is also exploring AI-controlled craniotomies, a significantly more complex task, while prioritising safety before moving on to live surgeries.
The Future of AI in Scientific Research
Santero is optimistic about the increasing collaboration between AI agents and human scientists, particularly as these agents gain direct access to laboratory equipment. He emphasises that human oversight remains crucial, especially in experiments involving live animals. It is still too early to consider allowing an AI agent to operate on a human brain.
As for the mouse involved in the experiment, Santero reports that it is recovering well, demonstrating the potential for integration of AI in medical research while highlighting the need for responsible and cautious application of such technology.