{"slug": "an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the", "title": "An AI agent ran part of a mouse brain surgery and refused to continue when the rig was miscalibrated", "summary": "Exin Therapeutics, a Y Combinator startup building AI agents for brain drug discovery, says one of its agents selected a gene therapy candidate and performed the injection stage of brain surgery on a mouse with epilepsy, then refused to continue when a calibration fault pushed its computed target outside the rig's operating range. Co-founder and CEO Gabriel Ocana-Santero described the run in an X thread on 28 October, saying the agent injected 250 nanolitres over five minutes while humans performed the craniotomy and a veterinarian held a physical emergency stop. The company says the mouse recovered, but whether the therapy reduced seizures is still being evaluated and the work is a research-stage demonstration only.", "body_md": "**🧠 An AI agent ran part of a mouse brain surgery and refused to continue when the rig was miscalibrated**\n\n📅 30 October 2026 • The Frontier\n\nExin Therapeutics, a Y Combinator startup building AI agents for brain drug discovery, says its agent ran an injection stage of brain surgery on a mouse with epilepsy. Co-founder and chief executive Gabriel Ocana-Santero described the run in a thread posted on X on 28 October: the team handed the agent a **spreadsheet of 19 gene therapy candidates**, the mouse epilepsy model, a custom motorised rig with cameras and motors exposed as tools, and one goal, to reduce seizures. The agent picked the treatment itself, worked out the target from the scientific literature, **lateral entorhinal cortex at 3.90 mm behind bregma, 4.00 right and 4.30 down**, then drove the arm and **injected 250 nanolitres over five minutes** while a camera watched the site for bleeding and backflow.\n\nThe run stopped on a safety check. After a disconnect, the computed Y-axis target of **34 mm exceeded the rig's 30 mm operating range**, and the agent refused to move, telling the operator **Nothing moved, the needle is still exactly at bregma zero, I am not retrying** and then That rules out a stale zero reference, I am not trying a third time. The instruction to continue anyway was overruled. The diagnosis proved correct: manual moves during an earlier pause had left the controller out of sync with the rig, and the code had to be fixed by hand while the animal was under anaesthesia with the needle in place.\n\nHumans held every irreversible step. A human surgeon performed the craniotomy, a step the agent is not allowed to attempt for safety reasons, and a licensed veterinarian approved every irreversible action and kept a physical emergency stop that could cut power to the rig's motors. On the models behind the run, The Frontier describes Anthropic's Opus family as the system used to assess surgical safety risk, while other coverage of the same thread reports that **Claude Opus 5.5 refused once it flagged the task as biological experimentation**, that Opus 5 completed a dry run with water and graph paper but declined the live procedure, and that Sonnet 5 carried out the injection.\n\nThe standing of the result matters as much as the demo. This is a **research-stage demonstration** that does not affect clinical practice or medical devices for humans, the mouse recovered, and whether the therapy actually reduced seizures is still being evaluated while peer-reviewed results are awaited. The primary account is the founder's own post on X, and no Anthropic comment on the experiment has been published.\n\n**IMPACT:** Teams putting agents in front of physical or regulated processes should copy this structure before scaling: computed targets validated against declared machine limits before every irreversible move, named human approvals with a physical kill switch, and a harness that lets the agent refuse an operator override. The refusal is what caught a real calibration fault here, and a system designed only to obey would have pushed a needle inside a brain on a desynchronised controller.\n\n**Continue the conversation:** [https://t.me/AIAutomationArchitecture](https://t.me/AIAutomationArchitecture) · [Bluesky](https://bsky.app/profile/delimitter.bsky.social)", "url": "https://wpnews.pro/news/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the", "canonical_source": "https://dev.to/delimitter/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the-rig-was-miscalibrated-44g7", "published_at": "2026-10-30 12:14:43+00:00", "updated_at": "2026-10-30 12:21:48.468543+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-startups", "artificial-intelligence"], "entities": ["Exin Therapeutics", "Gabriel Ocana-Santero", "Anthropic", "Claude Opus 5.5", "Claude Sonnet 5", "Y Combinator", "X"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the", "markdown": "https://wpnews.pro/news/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the.md", "text": "https://wpnews.pro/news/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the.txt", "jsonld": "https://wpnews.pro/news/an-ai-agent-ran-part-of-a-mouse-brain-surgery-and-refused-to-continue-when-the.jsonld"}}