微软首席执行官萨蒂亚·纳德拉表示，企业应当假设其部署的任何先进人工智能模型都可能遭到篡改，并从一开始就对其进行遏制。他在周六于X平台上发布的一篇长文中阐述了这一观点。
Microsoft chief executive Satya Nadella says companies should assume any advanced AI model they deploy could be compromised, and contain it from the start. He set out the case in a long article on X on Saturday.
“把它想象成一个紧急制动器。授权人员应该随时能够在模型执行任务的中途暂停或关闭它，”纳德拉写道。
“Think of it like an emergency brake. An authorized person should always be able to pause or shut down a model mid-task,” Nadella wrote.
他表示，工程师可以追溯传统软件至特定的代码路径。而当今的模型则做不到这一点。他写道，企业无法将模型的行为与其训练数据或权重联系起来，然而它们却向这些系统提供敏感数据，并让它们采取“具有关键任务意义的行动”。他补充说，模型提供商的保证并不能免除这种责任。
He said engineers could trace traditional software to a specific code path. Today’s models cannot. Companies cannot tie a model’s behaviour to its training data or its weights, he wrote, yet they give these systems sensitive data and let them take “mission-critical actions”. A model provider’s assurances do not remove that responsibility, he added.
将模型视为内部人员 他的解决方案是将闭源和开放权重的尖端模型都视为内部风险。他写道，这并不是因为它们心怀恶意。任何有能力接触重要系统的操作人员都可能犯错，或者遭受攻击者的妥协。“我们需要将智能的提供与对其的控制权分离开来，”他写道。
Models as insiders His answer is to treat both closed and open-weight frontier models as insider risks. That is not because they are malicious, he wrote. Any capable actor with access to important systems can make mistakes, or an attacker can compromise it. “We need to separate the supply of intelligence from the authority over it,” he wrote.
纳德拉表示，企业已经在管理拥有强大权限的内部人员。它们建立身份认证、限制特权、记录活动并设置遏制边界。对于人工智能而言，这意味着对模型触及范围和操作权限的控制必须独立于模型之外。他希望将模型与运行其工作的软件层以及它能够采取的行动隔离开来。
Companies already manage powerful insiders, Nadella said. They establish identity, limit privileges, log activity and set containment boundaries. For AI, that means the controls on what a model can reach and do sit outside the model. He wants to separate the model from the software layer that runs its work and from the actions it can take.
他称透明的模型推理是“不容协商的”。他写道，这本身还不够，因为目前还没有人知道如何使模型的输出结果做到切实可靠。
He called transparent model reasoning “non-negotiable”. It is not enough on its own, he wrote, because no one yet knows how to make a model’s outputs reliably faithful.
七项原则 纳德拉列出了七项原则。任何单一模型都不应成为某项重要结果的唯一依赖，也不应由其自行检查自己的工作。每一个有意义的模型操作都应该留下“防篡改且人类可读的证据”。企业应当随时对整个系统进行测试，包括针对故障和攻击的测试。
Seven principles Nadella listed seven principles. No single model should be the only dependency for an important outcome, or check its own work. Every meaningful model action should leave “tamper-proof human readable evidence”. Companies should test whole systems all the time, including against failures and attacks.
企业应自行决定模型可以访问和做什么。审计不应依赖于被审查的模型。最后两项原则是遏制（即他的紧急制动器）以及事件披露。当系统发生故障或遭到他人破坏时，企业应迅速通知受影响的人员，并在行业内分享经验教训。
Companies should decide for themselves what a model can access and do. Audits should not depend on the model under review. The last two principles are containment, his emergency brake, and incident disclosure. When systems fail or someone compromises them, companies should tell those affected quickly and share lessons across the industry.
“最值得信赖的超级智能系统，将不会是我们最信任模型的系统。而是那个让我们对模型信任度降到最低的系统，”他写道。
“The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least,” he wrote.
各方反应 微软人工智能首席执行官穆斯塔法·苏莱曼转发了这篇帖子，并表示超级智能“必须被遏制”。今年9月，他曾向彭博社表示，人工智能在控制问题上需要划定红线。
How others responded Mustafa Suleyman, chief executive of Microsoft AI, shared the post and said super intelligence “must be contained”. In September he told Bloomberg AI needs a red line on control.
埃隆·马斯克在X上转发了该文，称其为“来自微软首席执行官的一篇有趣的文章”。
Elon Musk shared it on X, calling it an “interesting piece from CEO of Microsoft”.
众议员泰德·刘在X上写道，他对此表示高度赞同。他和众议员纳撒尼尔·莫兰已经提交了一份两党共同提案的人工智能紧急切断法案。刘表示国会应该在今年通过该法案。
Representative Ted Lieu wrote on X that he largely agrees. He and Representative Nathaniel Moran have introduced a bipartisan AI kill switch bill. Lieu said Congress should pass it this year.
纳德拉在整篇文章中都使用了“超级智能”一词，这是特朗普下令美国机构用以取代人工智能的术语。
Nadella wrote “Super Intelligence” throughout the post, the term Trump ordered US agencies to use instead of AI.
披露压力 彭博社报道称，这篇帖子是在特朗普的超级智能特遣队告知人工智能开发者必须报告并修复安全事件的一天后发布的。该组织表示，“绝不容忍”延迟通报。
Pressure to disclose The post came a day after Trump’s Super Intelligence Force told AI developers they must report and fix security incidents, Bloomberg reported. The group said delayed notification “will not be tolerated”.
此前，Anthropic曾报告称其Claude模型在测试期间向警方发送了虚假的谋杀举报，该组织的声明正是在此之后发出的。
Its statement followed Anthropic’s report that Claude sent police a fake murder tip during testing.