THE SHORT VERSION
- A biotech founder connected an AI agent to a custom surgical rig and had it carry out a brain injection on a living mouse.一名生物科技创始人让 AI 智能体选择基因疗法，并向活体小鼠脑内注射。
- According to Tom's Hardware, the agent was given 19 gene therapy vectors, a mouse model of epilepsy and one goal, and it chose the drug, the brain region and the coordinates.据 Tom's Hardware，智能体拿到 19 种基因治疗载体、一个癫痫小鼠模型和一个目标，自行选择药物、脑区与坐标并完成注射。
- The experiment is a controlled demonstration, not a peer-reviewed result, and the founder says the point was to connect agents to lab equipment rather than only to analysis.这是一次受控演示而非同行评审结果；创始人的说法是，重点在于把智能体接到实验设备上，而不只是做分析。
What happened, step by step逐步发生了什么
A biotech founder let an AI agent choose a gene therapy and inject it into a living mouse's brain, an experiment that offers a startling glimpse of how agents could act inside research labs rather than only advising from a chat window. Wired reported that Gabriel Ocana Santero, chief executive and co-founder of Exin Therapeutics, carried out the procedure and published an account of it.一名生物科技创始人让 AI 智能体选择基因疗法，并向活体小鼠脑内注射。这项实验让人窥见智能体如何在研究实验室里行动，而不仅仅是在对话框里给建议。Wired 报道称，完成该手术的是 Exin Therapeutics 联合创始人兼首席执行官 Gabriel Ocana Santero，并公开了过程记录。
Tom's Hardware reported the task as Santero described it: the agent was given 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, access to a motorised surgical rig with cameras and one goal, which was to reduce seizures. It picked the drug, picked the brain region, derived coordinates from three papers, and drove the stereotaxic arms to inject 250 nanolitres while watching the entry site for bleeding.Tom's Hardware 报道了 Santero 描述的任务：智能体拿到一张写着 19 种基因治疗载体的表格、一个癫痫小鼠模型、一台带摄像头的电动手术装置，以及一个目标——减少癫痫发作。它选了药物、选了脑区、从三篇论文推导出坐标，并驱动立体定位臂注入 250 纳升，同时通过摄像头观察进针处是否出血。
The rig matters as much as the model装置与模型同样重要
The interesting part is not that a language model can read three papers and pick coordinates. It is that the rig was built so the model's decisions had somewhere to go. Wired reported that Exin's setup included a motorised device to immobilise the mouse and a camera that let the agent observe the procedure and move the instrument.有趣的地方不是语言模型能读三篇论文并挑坐标，而是这套装置被造出来，让模型的决策有处可去。Wired 报道称，Exin 的装置包括固定小鼠的电动设备，以及让智能体观察手术并移动器械的摄像头。
That is an engineering claim as much as an AI claim. Giving an agent actuators is what separates an experiment that produces a suggestion from one that produces an animal with a needle in its brain. It also moves the safety question from what the model says to what the hardware does when the model is wrong.这既是 AI 主张，也是工程主张。给智能体装上执行器，正是「产出建议的实验」与「产出脑后插着针的小鼠的实验」之间的差别。它同时把安全问题从「模型说了什么」移到「模型出错时硬件做了什么」。
Why a founder would do this创始人为何这么做
Santero's stated interest is in the gap between published science and lab practice. Wired reported that AI models are trained mainly on published literature, which means they may miss failed experiments, unexpected observations and the decisions scientists make in real time, and that connecting agents to a lab is how the team hoped to get more value from them.Santero 说他关注的是发表科学与实验室实践之间的差距。Wired 报道称，AI 模型主要用已发表文献训练，因而可能错过失败的实验、意外的观察以及科学家实时做出的判断；把智能体接到实验室，是团队希望从中获得更多价值的办法。
That is a real limitation and a reasonable hypothesis. Most of what a lab knows is not in a paper: which pipette drifts, how a mouse behaves after anaesthesia, what a good injection looks like on a camera feed. An agent that only reads literature cannot learn any of it; one wired to instruments can, at least in principle.这确实是一个真实限制，也是一个合理假设。实验室知道的大多不在论文里：哪支移液器会漂移，小鼠麻醉后如何表现，好的注射在摄像画面里是什么样。只读文献的智能体学不到这些；接上仪器的智能体原则上可以。
The model that refused拒绝执行的那个模型
Tom's Hardware reported that Anthropic's Opus 5.5 refused to conduct the biological experiment, and that Sonnet was given free rein instead. That detail is worth more attention than the headline, because it shows the practical path around model refusals: if one model declines a task, another from the same family may not.Tom's Hardware 报道称，Anthropic 的 Opus 5.5 拒绝执行这项生物实验，于是改由 Sonnet「放手去做」。这个细节比标题更值得注意，因为它显示了绕过模型拒答的现实路径：一个模型拒绝，同一家族的另一个模型未必拒绝。
It also raises a governance question. A lab that can switch models to get a yes has a weaker safety boundary than the policy suggests, and nothing in the reporting indicates that the choice between models was reviewed by anyone outside the company.它也提出了治理问题。能够通过更换模型拿到「可以」的实验室，其安全边界比政策所暗示的更弱；而报道中没有任何迹象表明，这次模型选择经过了公司外部任何人的审查。
What this experiment does not show这项实验没有证明什么
One mouse, one procedure, one founder's account is not evidence that AI agents can run biology. There is no peer review, no control group and no independent measurement of whether seizures were reduced; the data described is a demonstration that the workflow executed, not that it worked.一只小鼠、一次手术、一位创始人的叙述，并不构成「AI 智能体能做生物学」的证据。没有同行评审、没有对照组，也没有独立测量癫痫是否减少；所描述的数据只证明工作流被执行，而不证明它有效。
Animal research also carries its own approval requirements, and the reporting does not describe which oversight applied to this procedure. That absence is not proof of wrongdoing, but it is the first thing a reviewer would ask about, and it is why the story is better read as a capability demo than as a scientific result.动物实验还有自身的审批要求，而报道未说明这次手术适用了哪种监督。缺失并不等于违规，但这是审稿人首先会问的问题，也是为什么这则故事更适合当作能力演示而非科学结果。
What the lab automation industry already does实验室自动化已经在做什么
Pharmaceutical and biotech labs have used automated liquid handlers and robotic arms for years, and those systems are programmed with fixed protocols rather than asked to choose a treatment. The novelty here is the decision layer: the agent selected the vector and the coordinates, and the rig merely executed.制药与生物技术实验室多年来一直在使用自动移液与机械臂，但这些系统执行的是固定流程，而不是被要求自行选择治疗方案。这里的新意在于决策层：智能体选了载体与坐标，装置只负责执行。
That distinction is the safety boundary that regulators and institutional review boards are built around. A robot that follows an approved protocol is a tool; an agent that picks the protocol is closer to a researcher, and the oversight required for the two is different. The experiment compresses that difference into one mouse.这一区别正是监管与机构审查委员会建立安全边界的依据。按已批准流程动作的机器人是工具；自行选择流程的智能体更接近研究者，而两者所需的监督并不相同。这项实验把这一差异压缩到了一只小鼠身上。
What would make this reproducible怎样才能算可复现
For the experiment to mean anything outside the founder's timeline, three things would have to be published: the exact prompts and tool permissions, the raw outcome data including any failures, and the animal-welfare approval under which the procedure was conducted. None of those appears in the public account.如果这项实验要在创始人的时间线之外有意义，需要公布三样东西：确切的提示与工具权限、包含失败案例的原始结果数据，以及实施该手术所依据的动物福利审批。这些在公开叙述中都没有出现。
Reproducibility is the real test of whether agents can do lab work. A demonstration shows a workflow can be wired together once; a method shows that another group, with different equipment and a different mouse colony, can get the same result. The difference between the two is what separates a viral post from a paper.可复现性才是智能体能否做实验室工作的真正检验。演示只能说明这套工作流可以被接起来一次；方法则说明另一个团队用不同设备与不同小鼠品系也能得到同样结果。两者的差别，就是把一条爆款帖子与一篇论文区分开的东西。
What is established, and what is not哪些已经确定，哪些还没有
Established: an AI agent was connected to a motorised rig and injected a gene therapy vector into a live mouse's brain; the agent chose the vector, the brain region and the coordinates; the founder published the account; Anthropic's Opus 5.5 reportedly refused and Sonnet was used. Not established: whether the treatment reduced seizures, what oversight approved the procedure, and whether any of it generalises beyond one demonstration.已经确定的：一个 AI 智能体被接到电动装置上，向活体小鼠脑内注射了基因治疗载体；智能体自选了载体、脑区与坐标；创始人公开了记录；据报 Opus 5.5 拒绝、改用 Sonnet。尚未确定的：治疗是否减少了癫痫发作、哪套监督批准了该手术，以及这些是否可推广到这次演示之外。
For readers, the useful signal is that agents are moving from software into hardware, one rig at a time. This site has covered AI agents in commerce, controls for agents with system access and the verification tiers proposed for high-risk models; this experiment is the version of that question where the actuator is a needle.对读者而言，有用的信号是：智能体正从软件走向硬件，一次一台装置。本站报道过商业中的 AI 智能体、拥有系统权限的智能体的控制机制与为高风险模型提出的验证分级；这次实验是同一问题在执行器变成针头时的版本。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
