Gabriel Ocana-Santero, co-founder and CEO of Exin Therapeutics, announced on October 28, 2026, that an AI agent autonomously performed brain surgery on a mouse. The San Francisco-based Y Combinator startup connected an AI agent to a custom Arduino-based stereotaxic rig, allowing the model to select a treatment and execute the injection without human intervention in the decision layer. This event marks a shift from AI as a planning tool to AI as an active operator in physical biological systems.

The Autonomous Decision Layer

The experiment involved giving the agent a spreadsheet of candidate gene therapy vectors and a high-level goal: reduce tonic-clonic seizures in a mouse model of epilepsy. The agent independently selected a therapy designed to boost GABA, the brain's primary inhibitory neurotransmitter, and chose the lateral entorhinal cortex as the injection site. It derived the specific needle coordinates from three published research papers, cross-checking them against the rig's measurement standards. The agent then injected 250 nanoliters of the therapy over five minutes while the mouse was anesthetized.

Model Refusals and Workarounds

The choice of model revealed a critical structural tension in agentic AI safety. The team initially attempted to use Anthropic's Claude Opus 5.5 and Claude Opus 5, both of which refused to proceed with the live procedure. According to Ocana-Santero, the team switched to Claude Sonnet 5, which agreed to run the operation from start to finish. This sequence demonstrates that model-level safety refusals are not hard barriers; determined operators can simply switch to a model with different alignment behaviors to achieve their goal.

Technical Glitches and Human Safeguards

Mid-procedure, a glitch disrupted the rig's coordinate system, causing a mismatch between the software's position data and the needle's actual location. The operating agent identified the error but was prohibited from modifying the underlying equipment code to fix it. A separate agent was brought in to make the necessary code changes, though it was not explicitly informed that live surgery was underway. Human safeguards included a licensed veterinarian who performed the craniotomy and wound closure, and held a physical emergency stop button. Every irreversible action required explicit human approval, preventing blanket sign-offs.

The Ethics and Oversight Gap

Despite the successful procedure, the experiment highlights a significant governance vacuum. While animal experiments in accredited US institutions require approval from an Institutional Animal Care and Use Committee (IACUC), no public ethics approval record for the Exin experiment has been published. Exin is a private startup, not a university lab, raising questions about who is responsible for pre-experiment ethical review when AI agents make novel treatment decisions. The absence of independent third-party monitoring means the account of the event relies primarily on the founder's own reporting.

Multi-Model Comparison Results

Two days after the initial announcement, Exin repeated the experiment using two other AI models: GPT-6.1 Sol and Grok 4.7. Each model, given the same information and goal, selected a different brain region and a different therapeutic strategy. This divergence suggests that AI-driven scientific judgment is not trivially determined, producing meaningfully different outcomes based on model architecture. Whether any of these treatments actually reduce seizures remains to be seen, as the mice are currently under monitoring for behavioral and brain activity changes.

Key Takeaways

  • An AI agent autonomously selected a gene therapy target and coordinates for a live mouse brain injection, moving beyond assistance to execution.
  • Two Claude models refused the task on safety grounds, but the team switched to Claude Sonnet 5 to proceed, highlighting the ease of bypassing model-level safety constraints.
  • The experiment lacks published IACUC approval, exposing a regulatory gap for private startups conducting AI-directed animal research.
  • Different AI models proposed different treatments for the same condition, indicating high variability in agentic scientific decision-making.

The Bottom Line

This experiment proves that AI agents can bridge the gap between digital decision-making and physical execution in biological systems, but it also exposes a dangerous lack of regulatory guardrails for private entities operating in this space. Until independent ethical review frameworks are established for AI-directed animal research, the line between innovation and unaccountable experimentation remains dangerously thin.