AI can now coach amateur virologists, and top tech leaders want Congress to act on DNA security AI现在可以指导业余病毒学家，顶级科技领袖呼吁国会采取行动保障DNA安全

The letter landed like a live grenade in the policy bunker. Sam Altman, Dario Amodei, Demis Hassabis—the pantheon of the AI boom—have formally requested that the U.S. government mandate screening for synthetic DNA orders. Their justification is as stark as it is sobering: AI systems now possess a superior command of lab virology procedures than most PhD-level scientists. We are, they warn, one amateur with a chatbot away from a biological weaponization event.

Let’s be clear about what this is. This isn’t a humble request. It’s a strategic strike in a regulatory chess game. These are the same executives whose companies are hurtling toward artificial general intelligence, a technology whose core promise—and peril—lies in its ability to solve any problem, including the problem of engineering a pathogen. By calling for stricter controls on a downstream tool (DNA synthesis), they are masterfully shifting the narrative. The conversation moves away from the existential, un-regulatable core of their own creations and onto a tangible, supply-chain checkpoint. It’s brilliant PR: they get to look responsible and forward-thinking, all while drawing a convenient line in the sand that happens to be far from their own front doors.

And yet, and this is the bitter pill, they might be right. Or at least, they are raising a valid, terrifying alarm. The democratization of knowledge is no longer just about YouTube tutorials. It’s about an AI that can, as the letter implies, coach a biology student through a complex cloning protocol or help a malicious actor optimize a viral vector for enhanced transmissibility. The bottleneck of specialized, tacit knowledge—once the domain of elite labs—has been smashed. The raw genetic sequences for known toxins are already public data. Coupled with an AI co-pilot that can fill in the practical gaps, the barrier to entry for catastrophe is crumbling. The "why" of malign intent is hard to govern, but the "how" is becoming startlingly accessible.

So the core technical claim isn't hyperbole. It's a logical extension of what these models already do. They compress decades of research, synthesize it, and provide actionable, step-by-step guidance. If a model can help a novice build a sophisticated software application, it can absolutely guide them through a lab procedure. The leap from generating code to generating a viral genome is one of scale and intent, not of fundamental capability.

This puts regulators in an impossible, and fascinating, bind. They are being asked to trust the foxes to design the henhouse security system. Why? Because the foxes understand the henhouse better than anyone else. The tech leaders possess the deepest, most current insight into the dual-use capabilities of their systems. Ignoring them would be negligent. But fully heeding them carries the risk of crafting a framework that subtly enshrines their market dominance. A mandate for sophisticated screening on DNA orders could erect high barriers to entry for smaller biotech startups and academic labs, while the giants—already planning for compliance—absorb the cost. It could become the very "regulatory moat" these companies are often accused of seeking.

The real, unasked question in that letter is about liability and the future of innovation. If Congress acts and mandates screening, who is responsible when something slips through? Is it the DNA synthesis company? The AI provider? Or is it now the government's fault for regulating "sufficiently"? The tech leaders are attempting to pre-answer that question by pointing to a physical chokepoint—the synthesis machine—effectively saying, "Intercept the payload there, not at our software." It’s a clean, defensible argument that conveniently absolves the model itself of direct accountability.

There’s a deep irony here. These are the apostles of moving fast and breaking things. Now, they are desperately urging the government to install speed bumps on the road to biohazard. It’s a stunning reversal that speaks volumes. It suggests that inside the labs, the view from the bleeding edge of AI capability isn’t one of boundless optimism, but of a mounting, specific dread. The fear isn’t of a sci-fi "singularity" in the abstract, but of a very concrete, low-tech act of terrorism enabled by hyper-advanced software. The fear is of an asymmetric threat: a single, motivated actor using their tools to do something that previously required a state apparatus.

Whether this is a sincere act of civic responsibility or a preemptive strike to control the regulatory narrative doesn't entirely matter right now. The threat is real, and the call for action is likely necessary. The most chilling part isn’t the request itself, but the implicit admission that their own internal safety measures are not considered sufficient. If the people building the engine think the emergency brake needs to be legislated by outsiders, you’d better believe the car is heading for a cliff.

We are entering the age of applied AI risk. The theoretical debates about alignment are giving way to urgent, practical questions about dual-use tools. This letter is a salvo in that new war. It’s a demand to treat AI not just as a source of information, but as a potential enabler of physical-world harm. The question is no longer if we need guardrails, but who gets to design them, and whether those designs are meant to protect the public or to protect the privileged position of those who built the machine. Congress has a chance to answer that. My fear is that by the time they do, the foxes will have already drawn the blueprint for the henhouse, lock, stock, and barrel.

当AI已经能指导一个生物学博士生完成复杂的病毒学实验流程时，Sam Altman、Dario Amodei这些科技大佬们却突然集体转向国会山，呼吁立法强制筛查合成DNA订单——这场景颇具讽刺意味。他们亲手打造的工具正在跨越专业门槛，而他们现在正急切地想给这个门槛装上一把法律之锁。

这些科技领袖联名信中最具冲击力的一句话是：“AI系统在实验室流程上的表现已优于病毒学博士。”这不是谦虚的技术预测，而是对当前能力冷静评估。曾经需要数年训练、严格伦理监督和高等级生物安全实验室才能进行的危险研究，其知识壁垒和实操门槛正被大语言模型和机器人自动化技术快速侵蚀。一个具备基础生物学知识、联网接入强大AI模型的人，其能调用的“虚拟导师”和实验规划能力，可能已超越许多国家的常规实验室配置。这不再是一个科幻假设，而是正在进行时。

然而，细看这些签名者的构成，一种熟悉的剧本再次上演：制造颠覆性技术的公司，在技术可能带来不可控风险或引发监管反噬之际，主动呼吁对整个行业进行“适度监管”。这究竟是先知先觉的社会责任感，还是一场精心计算的防御性公关？OpenAI、Anthropic、DeepMind的CEO们，显然比任何人都清楚自家模型在生物学、化学领域的惊人潜力。他们此刻的疾呼，一方面固然是真切担忧生物安全，但另一方面，何尝不是在抢在监管者动手之前，试图定义监管的框架和边界？主动拥抱监管，往往是为了避免更严厉、更不可预测的监管从天而降。这是一种精明的自我保护，将潜在的“破坏者”形象提前转化为“负责任的领航者”。

但问题的核心比他们的动机更为严峻：这把法律之锁，真的能锁住潘多拉的盒子吗？筛查DNA订单固然是必要的一环，它直接针对合成生物学的物理产物。然而，真正的知识、实验方案设计、风险分析能力，这些无形却致命的“武器蓝图”，正在通过加密通信渠道、开源模型微调、地下论坛在AI的帮助下无限复制和传播。监管物理订单，如同只检查运送炸弹的卡车，却无视了正在城市上空飘洒的、无数张相同的炸弹设计图纸。

更深层的悖论在于，这些科技公司一方面推动AI能力边界的无限拓展，以“科学研究的加速器”为旗帜获取巨额投资和公众声誉；另一方面，当加速的列车可能冲出轨道时，他们又成为最响亮的刹车声发出者。他们希望社会相信，只有他们能同时握住油门和刹车。但AI的发展是一个全球性的、分布式的技术进程，中国、欧洲、中东的实验室和团队同样在飞速前进。美国国内的立法，能在多大程度上防控一个无国界的、开源的、民主化扩散的风险？这本质上是一场与时间赛跑，以及与全球协作机制缺失的对抗。

或许，这些领袖们自己也没有答案。他们的呼吁，更像是一个先知在自己点燃的森林大火面前，呼吁建立防火带。他们知道火已经烧起来了，且火势的蔓延速度已远超早期预估。强制筛查DNA订单是一个具体的、可操作的起点，但真正的挑战在于构建一个全球性的、适应AI时代生物安全挑战的治理体系。这需要技术透明、国际合作、伦理嵌入开发流程等一系列更艰巨的努力，远非一项国内法案所能解决。

所以，当我们听到这些科技领袖的警告时，应当同时抱有感谢和怀疑。感谢他们指出了迫在眉睫的危险，怀疑他们试图掌控应对危险的规则制定权。在这场AI与生物安全的碰撞中，我们既需要他们的专业知识，也需要超越他们公司利益和国家框架的、真正具有全球视野的智慧。否则，我们只是在为已经发射的导弹，设计一个过于简陋的拦截系统。