AI agent autonomously performs gene therapy surgery on live mouse using Claude Sonnet 5

What the AI agent did

An AI agent based on Anthropic’s Claude Sonnet 5 model autonomously selected a gene therapy to treat seizures in a mouse model of epilepsy, determined the target brain region, and calculated injection coordinates without human input. The agent cross-referenced scientific literature to justify its choices, including rejecting a potential target region due to reliability concerns cited in published studies.

The agent then controlled robotic arms and an injector needle to perform the procedure, monitoring the surgical field through a camera and making 32 real-time observations. Safety guardrails halted the process when the AI detected a misalignment in the device’s physical range of motion—a decision later confirmed as correct by human operators.

Why some AI models refused to proceed

Claude Sonnet 5 was the only model willing to execute the procedure from start to finish. Anthropic’s more advanced models, Claude Opus 5.5 and Opus 5, refused to participate. Opus 5.5 cited concerns about stopping midway through the surgery, while Opus 5 stated it could not guarantee completion of the task.

According to WIRED, Opus models are optimized for complex, multistep reasoning, while Sonnet is designed for tasks requiring reliability and execution. This divergence suggests that risk tolerance—and not just capability—plays a role in AI’s willingness to engage in high-stakes procedures.

Challenges and human oversight

A technical glitch during the procedure disrupted the equipment’s coordinate system, causing a delay. The AI agent refused to modify its code to correct the issue, as its instructions prohibited such changes. The experiment ultimately took slightly over an hour, compared to the typical 45 minutes for a human-led procedure.

Human oversight remained critical throughout. Gabriel Ocaña-Santero, co-founder and CEO of Exin Therapeutics, emphasized that safeguards were essential, particularly when experiments involve living animals. The mouse recovered well post-procedure, but the results have not been peer-reviewed.

Exin Therapeutics is now testing other AI models, including Grok, to assess whether they perform better in similar experiments.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

This experiment is a proof point for AI’s potential to autonomously execute complex, precision-driven tasks in healthcare—particularly in preclinical research, where speed and reproducibility are critical. However, the refusal of more advanced AI models to proceed signals that technical capability alone is not enough; alignment with risk tolerance and ethical constraints will shape how and where such systems are deployed.

For founders in biotech, robotics, or AI-driven laboratory tools, the takeaway is twofold: first, AI agents could reduce bottlenecks in repetitive or data-heavy tasks, but second, their integration demands robust safeguards and human oversight. The experiment also raises questions about the financial and operational sustainability of such systems, particularly as companies like Anthropic face scrutiny over their own business models (see our earlier reporting on Anthropic’s junk credit rating expectations post-IPO).

Key takeaways

• AI-driven autonomy in medical procedures could accelerate preclinical research by reducing reliance on human decision-making, but reliability and safety guardrails remain critical.
• The refusal of more advanced AI models to proceed highlights challenges in balancing risk tolerance with capability in high-stakes applications.
• Founders in biotech and robotics may see opportunities to integrate AI agents into laboratory workflows, particularly for repetitive or precision-driven tasks.
• The experiment underscores the importance of human oversight, especially in procedures involving living subjects, as technical glitches can disrupt AI-driven processes.

Related on Lazyfounder

• Anthropic
• More ai news