Trump administration officials have asked a government artificial-intelligence testing unit to stop issuing public reports, the latest signal that the White House is tightening control over AI models as national-security concerns increase.

特朗普政府官员已要求一个政府人工智能测试机构停止发布公开报告，这是随着国家安全担忧加剧，白宫正收紧对人工智能模型管控的最新信号。

Administration officials including National Cyber Director Sean Cairncross have told the Center for AI Standards and Innovation to halt publication of its model assessments while an executive order President Trump signed last week is implemented, people familiar with the matter said. The order represented a win for Cairncross and Treasury Secretary Scott Bessent, who have pushed for security considerations to play a bigger role in model evaluation.

知情人士称，包括国家网络主任肖恩·凯恩克罗斯在内的行政官员已告知“人工智能标准与创新中心”，暂停发布其模型评估报告，以便实施特朗普总统上周签署的一项行政命令。 该命令的出台代表着凯恩克罗斯和财政部长斯科特·贝森特的胜利，他们一直力主安全考量在模型评估中发挥更大作用。

The concern was prompted by the recent release of Anthropic's Mythos and other powerful models capable of carrying out cyberattacks or potentially aiding the creation of biological weapons. The administration is working with Anthropic and OpenAI to control who has access to their best products to make sure the companies are protecting national security.

此举的导火索是近期Anthropic公司的“Mythos”及其他能够实施网络攻击或可能协助制造生物武器的强大模型的发布。 政府正在与Anthropic和OpenAI合作，控制谁能接触到它们最好的产品，以确保这些公司保护国家安全。

The move to halt the center's public work is being viewed as a sign to some officials that Cairncross and his allies want more say over model evaluation, the people said. Some administration officials are upset at Cairncross for exerting more influence over the process; they thought the executive order directed a new group to do work that the center—known as CAISI—was already doing, the people said.

知情人士称，暂停该中心公开工作的举动，在部分官员看来，表明凯恩克罗斯及其盟友希望在模型评估上拥有更大发言权。 一些政府官员对凯恩克罗斯在此过程中施加更多影响力感到不满；这些人表示，他们认为该行政命令是指导一个新机构去完成该中心（即CAISI）已经在做的工作。

On the other side are companies including OpenAI, which have had discussions with administration officials about the importance of the center, known as CAISI, and preserving its power, the people said. White House AI advisers including venture capitalist David Sacks have warned that an overzealous model-testing process—regardless of who oversees it—could slow deployment and hinder innovation.

知情人士称，另一方面则是包括OpenAI在内的公司，它们已与政府官员讨论了该中心（即CAISI）的重要性以及保留其权力的必要性。 包括风险投资家大卫·萨克斯在内的白宫人工智能顾问警告称，无论由谁监督，一个过于热心的模型测试过程都可能延缓部署并阻碍创新。

"The implementation of President Trump's AI agenda is a whole-of-government effort, with numerous agencies contributing to its success. The entire administration is working closely together," White House spokeswoman Liz Huston said. Cairncross is doing excellent work to protect the American people while also promoting innovation, she said.

白宫发言人莉兹·休斯顿表示：“特朗普总统人工智能议程的实施是一项‘整体政府’的努力，众多机构都为它的成功做出了贡献。 整个政府正在紧密合作。” 她表示，凯恩克罗斯在促进创新的同时，也为保护美国人民做了出色的工作。

Housed within the Commerce Department, CAISI is the main body within the government that tests AI models before they are released and puts out public information about their capabilities and relative performance. It is seen as one of the primary vehicles by which the government can mitigate risks tied to cybersecurity, biological weapons and other AI threats.

CAISI隶属于商务部，是政府在人工智能模型发布前对其进行测试，并发布有关其能力和相对性能的公开信息的主要机构。 它被视为政府可以用来减轻与网络安全、生物武器和其他人工智能威胁相关风险的主要工具之一。

CAISI is still working internally to evaluate models and coordinate with government agencies, but stopping its public work has thrown its future in jeopardy, the people said.

知情人士称，CAISI仍在内部开展模型评估并与政府机构协调，但停止其公开工作已使其未来陷入不确定之中。

Top model developers including OpenAI and Anthropic have relationships with CAISI going back to the Biden administration. The process described in the executive order could challenge how companies plan to release models if they are concerned that national-security officials might deem them risky and slow down their deployment, AI analysts said.

包括OpenAI和Anthropic在内的顶尖模型开发商与CAISI的关系可以追溯到拜登政府时期。 人工智能分析师表示，如果公司担心国家安全官员可能认为其模型有风险并减缓其部署，那么行政命令中描述的过程可能会对它们计划发布模型的方式构成挑战。

The move to restrict CAISI's activity is the latest example of how powerful models such as Mythos have thrown the administration's hands-off AI strategy into disarray.

限制CAISI活动的举措，是像“Mythos”这样的强大模型如何让政府“不干预”的人工智能策略陷入混乱的最新例证。

As a civilian agency, CAISI was previously able to share findings with the public and other researchers, a process analysts say is unlikely to continue if security agencies have more control.

作为一个民事机构，CAISI此前能够与公众和其他研究人员分享研究成果，分析师表示，如果安全机构获得更多控制权，这一过程不太可能继续。

In a national-security memo published by the White House last week, a section about AI threats mentions security officials and excludes CAISI.

在白宫上周发布的一份国家安全备忘录中，关于人工智能威胁的部分提到了安全官员，却未提及CAISI。

White House officials instructed CAISI to remove an announcement last month saying that it would work with Microsoft, Google and Elon Musk's AI company while the executive order was completed, people familiar with the matter said.

知情人士称，白宫官员曾指示CAISI撤下上个月的一则公告，该公告称，在行政命令完成期间，该中心将与微软、谷歌和埃隆·马斯克的人工智能公司合作。

In April, a former Anthropic researcher was days into leading CAISI when the White House abruptly asked him to resign, a move some officials felt stemmed from the administration's feud with the AI developer.

今年4月，一名前Anthropic研究人员在领导CAISI仅几天后，就被白宫突然要求辞职，一些官员认为此举源于政府与该人工智能开发商之间的不和。

Commerce Secretary Howard Lutnick reimagined CAISI at the start of Trump's second term, changing its name from the AI Safety Institute, the original name under former President Joe Biden.

商务部长霍华德·卢特尼克在特朗普第二任期开始时对CAISI进行了重新定位，将其从拜登总统任内的原名“人工智能安全研究所”更改过来。

Last week, OpenAI called for CAISI to be strengthened with more resources and responsibilities in a policy blueprint. Chris Lehane, the company's chief global affairs officer, told reporters that the unit was capable of sophisticated testing and had relationships with leading companies.

上周，OpenAI在一份政策蓝图中呼吁，应通过增加资源和职责来加强CAISI。 该公司首席全球事务官克里斯·莱恩告诉记者，该机构有能力进行复杂的测试，并与领先公司建立了关系。

AI experts say the unit is underfinanced compared with similar groups in other countries and is needed to provide U.S. leadership in the field.

人工智能专家表示，与其他国家类似机构相比，该机构资金不足，但对于美国在该领域发挥领导作用而言是必需的。