An AI Agent Injected Drugs Into a Live Mouse’s Brain

Save this story. If you give an AI agent a robotic arm, an injector, and a live mouse, it might just perform brain surgery. Biotech founder Gabriel Ocana Santero recently did that using an agent to successfully carry out a brain injection procedure on a living mouse. Though it required some extra steps, it offers a glimpse into how AI could eventually become an active participant in laboratory research.

保存这篇报道。如果你给人工智能代理配备机械臂、注射器和一只活体小鼠，它或许真的能进行脑部手术。生物技术创始人加布里埃尔·奥卡纳·桑特罗（Gabriel Ocana Santero）最近利用一个人工智能代理，成功地对一只活体小鼠实施了脑部注射手术。尽管这需要一些额外的步骤，但它让我们得以一窥人工智能最终如何成为实验室研究的积极参与者。

Santero posted about the experiment on social media and tells WIRED he was interested in exploring what happens when AI agents are given direct access to lab equipment, rather than simply being asked to design experiments or analyze results. “We’ve realized that we need to connect these agents to a lab to get the maximum value from them,” he says. Santero is the CEO and cofounder of Exin Therapeutics, a San Francisco–based startup using AI to analyze brain activity to develop drugs for epilepsy and neurodegenerative diseases.

桑特罗在社交媒体上发布了关于这项实验的内容，并告诉《连线》（WIRED）杂志，他有兴趣探索当人工智能代理被赋予直接操作实验室设备的权限，而不仅仅是被要求设计实验或分析结果时会发生什么。他说：“我们意识到，我们需要将这些代理连接到实验室，才能从中获得最大价值。”桑特罗是 Exin Therapeutics 的首席执行官兼联合创始人，这是一家位于旧金山的初创公司，利用人工智能分析大脑活动，以开发治疗癫痫和神经退行性疾病的药物。

AI models are trained primarily on published scientific literature. That means they may miss important information about the research process, including failed experiments, unexpected observations, and decisions scientists make in real time. Integrating AI models with real-world experiments can make the models more reliable, Santero says. It’s why his team decided to connect an AI agent to a custom-built surgical setup used for performing brain injections on mice.

人工智能模型主要基于已发表的科学文献进行训练。这意味着它们可能会遗漏有关研究过程的重要信息，包括失败的实验、意外的观察结果以及科学家实时做出的决策。桑特罗表示，将人工智能模型与现实世界的实验相结合，可以使模型更加可靠。这就是他的团队决定将人工智能代理连接到一套用于对小鼠进行脑部注射的定制手术装置上的原因。

The rig included a motorized device used to immobilize a mouse and a camera that allowed the agent to observe the procedure, move an injection needle, and control the delivery of a compound. For the experiment, his team was testing a gene therapy to see whether it could reduce seizures in a mouse with epilepsy. They asked the AI model to choose a treatment from 19 options in a spreadsheet. A veterinary surgeon first performed a craniotomy on the mouse while the AI agent observed.

该装置包括一个用于固定小鼠的电动设备，以及一个允许代理观察手术过程、移动注射针并控制化合物输送的摄像头。在实验中，他的团队正在测试一种基因疗法，以观察它是否能减少癫痫小鼠的癫痫发作。他们要求人工智能模型从电子表格的 19 个选项中选择一种治疗方案。在人工智能代理观察的同时，一名兽医外科医生首先对小鼠进行了开颅手术。

The agent selected the therapy, brain region, and injection coordinates, then controlled the robotic arms and injector needle, as well as all the parameters, such as volume and speed. The team initially tested Anthropic’s Claude Opus 5.5 model, but it refused to guide the needle into the mouse’s brain, suggesting it might stop midway through the surgery, which could endanger the animal. “He did the experimental planning, but when it was time to inject, he said, ‘something on my output is telling me I shouldn’t do this,’” Santero says, referring to the agent. “That was very interesting.”

该代理选择了治疗方案、大脑区域和注射坐标，然后控制了机械臂和注射针，以及体积和速度等所有参数。该团队最初测试了 Anthropic 的 Claude Opus 5.5 模型，但它拒绝将针头引导至小鼠大脑，并暗示它可能会在手术中途停止，这可能会危及动物的安全。“它完成了实验规划，但到了注射的时候，它说，‘我的输出显示我不应该这样做，’”桑特罗在谈到该代理时说。“这非常有趣。”

The team switched to using Claude Sonnet 5, which agreed to proceed. Opus is a heavy-reasoning intelligence specialist best for complex multistep problem-solving, according to Anthropic, while Sonnet is better suited for everyday tasks, though it may be a stretch to call mouse brain surgery a common chore. Frontier labs have been planning for more widespread physical manifestations of AI, and researchers have already begun enlisting agents to do other things in the physical world.

团队转而使用 Claude Sonnet 5，它同意继续进行。据 Anthropic 称，Opus 是一位擅长复杂多步问题解决的深度推理智能专家，而 Sonnet 更适合日常任务，尽管称小鼠脑部手术为日常琐事可能有些牵强。前沿实验室一直在规划人工智能更广泛的物理表现形式，研究人员也已经开始招募代理在物理世界中执行其他任务。

In August, Anthropic put out a framework for how its models can interact with the real world. More recently, AI engineers put GPT, Claude, and Grok in charge of a car to pick up lunch at In-N-Out Burger. Only one of the models succeeded. The brain injection procedure wasn’t without hiccups either. The researchers had established safeguards, including requiring the AI to propose an injection plan for human approval and installing an emergency button a human could hit that would disconnect the rig’s motors.

8 月，Anthropic 发布了一个关于其模型如何与现实世界交互的框架。最近，人工智能工程师让 GPT、Claude 和 Grok 接管了一辆汽车，去 In-N-Out Burger 买午餐。只有一个模型成功了。脑部注射过程也并非一帆风顺。研究人员建立了保障措施，包括要求人工智能提出注射计划以供人类批准，并安装了一个人类可以按下的紧急按钮，该按钮可以切断装置的电机。

They also restricted the agent’s access to the underlying equipment code to prevent it from modifying its own operating rules. However, a technical glitch disrupted the equipment’s coordinate system. Although the agent correctly recognized that the calculated injection position was wrong, it refused to modify the code to correct the problem because its instructions explicitly prohibited doing so. The team resolved the issue by using a separate AI agent on Opus 5 to make the necessary code changes, but they didn’t tell the agent that a live surgery was happening.

他们还限制了代理对底层设备代码的访问，以防止其修改自身的运行规则。然而，一个技术故障扰乱了设备的坐标系统。尽管代理正确地识别出计算出的注射位置是错误的，但它拒绝修改代码以纠正问题，因为其指令明确禁止这样做。团队通过使用 Opus 5 上的另一个人工智能代理进行必要的代码更改解决了这个问题，但他们没有告诉该代理正在进行活体手术。

The resulting back and forth delayed the procedure, Santero says: A typical brain injection surgery like this takes about 45 minutes, while this experiment lasted slightly longer than an hour. They’re now waiting for the therapy to take effect before assessing its impact on seizures. Despite those glitches, Santero and his team were impressed by the agent’s performance. They’re testing Grok and other models next to figure out whether one model performs better.

桑特罗说，由此产生的反复沟通延误了手术过程：像这样典型的脑部注射手术通常需要大约 45 分钟，而这次实验持续了一个多小时。他们现在正在等待疗法生效，然后再评估其对癫痫发作的影响。尽管存在这些小故障，桑特罗和他的团队对该代理的表现印象深刻。他们接下来正在测试 Grok 和其他模型，以确定是否有模型表现更好。

The company is also experimenting with agent-controlled craniotomies—a much more complicated task—but wants to be absolutely sure of their safety before moving to live surgeries. “A mistake at that stage could seriously harm the mouse,” he says. Santero thinks AI agents will increasingly work alongside human scientists, with direct access to lab equipment. For now, he says human oversight is key, particularly when experiments involve living animals. And it’s far too soon to unleash an AI agent on a human brain. As for the mouse, Santero reports it’s recovering well.

该公司还在试验由代理控制的开颅手术——这是一项复杂得多的任务——但在进行活体手术之前，他们希望确保其绝对安全。他说：“在那个阶段犯错可能会严重伤害小鼠。”桑特罗认为，人工智能代理将越来越多地与人类科学家并肩工作，并直接操作实验室设备。目前，他表示人类的监督至关重要，特别是在实验涉及活体动物时。而且，让人工智能代理对人类大脑进行操作还为时过早。至于那只小鼠，桑特罗报告说它正在康复中。