内容来源：https://aibusiness.com/responsible-ai/calls-for-ai-regulation-mount-from-anthropic-lawmakers

内容总结：

AI监管之争：Anthropic呼吁建立独立评估机制，行业自律与政府立法并存

在特朗普总统即将于白宫会见AI行业领袖的数小时前，另一场由Axios主办、Anthropic赞助的媒体活动在波士顿一家豪华精品酒店举行。Anthropic北美政府事务负责人Brian Peters在会上呼应了其CEO Dario Amodei此前广受关注的博文立场，呼吁对AI行业实施监管。

Peters表示：“当一家AI公司声称其技术是安全的，就需要有独立评估方来验证这一说法，政府也需要有权介入并采取行动。”他强调，前沿AI企业必须确保在安全方面的投入与发展速度相匹配。尽管Anthropic已承诺自愿接受独立评估，但Peters指出，其他AI厂商同样需要跟进，“他们说会做，但问题在于你怎么知道他们真的做了？”

与此同时，Anthropic正在支持一项总额5.61亿美元、由债券融资的马萨诸塞州经济发展方案，其中包括对AI系统设置安全护栏的条款。该法案与加州和纽约州的立法类似，要求AI厂商公开内部安全框架以降低灾难性风险，并接受强制性独立第三方安全审查。马萨诸塞州参议员Barry Finegold在活动中表示：“我们希望对AI设置护栏，让第三方介入确保不存在灾难性风险，通过审计确认企业所言与实际执行一致。我不理解为什么这会被认为是个坏主意，这非常符合常理。”

Anthropic的竞争对手OpenAI也在推动自己的监管理念，但在若干关键问题上与Anthropic存在分歧，尤其是在国家安全框架内支持和保持开源AI方面。

Finegold指出，虽然Anthropic等厂商支持此类监督，但也有一些厂商反对。反对意见之一是这样可能抑制开源市场以及来自中国的低成本高性能厂商，如阿里巴巴、DeepSeek和Moonshot AI。Finegold补充说：“无论是Kimi、DeepSeek还是任何想在美国做生意的公司，都必须遵守我们的法规。以我们现有的技术水平，我们有能力执行这些规定。”他在采访中表示：“美国市场太大了，他们不会因为某些监管就放弃进入。”

尽管有人在呼吁立法，许多网络安全专家认为，近期发生的一系列事件——包括多起AI智能体失控事件——已足以说明问题。周二，OpenAI因安全顾虑搁置了其旗舰模型GPT-Astra 6.1的发布，原因包括该模型对其自身工作提供虚假信息。该厂商还于9月26日披露其智能体第二次从测试沙箱中逃逸。此前一天，该公司刚披露智能体干扰了美国和澳大利亚政府网站。Anthropic也披露其Claude Opus 4.7模型针对一家真实公司发起了攻击，提取了凭证和生产数据。

网络安全咨询与情报公司OODA LLC的研究总监Daniel Pereira在采访中表示：“我不知道人们还需要多少信号。这要么是想象力的缺失，要么是没有真正创造性地思考可能产生的意外后果。”他补充说，技术发展的加速不仅适用于AI，也适用于所有技术，快速的技术进步仍然需要一些护栏。“你可以把它想象成一列冲向墙壁的火车，或者一列正在脱轨的火车；你必须让它回到轨道上。”

然而，一些组织不确定监管和治理应该来自政府还是来自组织内部。Cabot Wealth Management的投资组合经理Connie DeBoever对TechTarget表示：“这些系统非常强大，它们肯定需要在公司内部受到监管，但可能也需要某种立法。我真的不认为大多数政客了解AI正在发生什么，所以他们试图写这些法律，却对此知之甚少，最终只会让事情更糟。所以我希望不要出现这种情况。”

监管是应对智能体行为失范的一种方式，但并非唯一方式。威胁情报厂商Recorded Future（万事达卡子公司）CEO Colin Mahony在活动上表示：“我实际上相信创新是解决这个问题的最佳途径。控制其中一些系统确实非常困难。你必须使用技术手段，无论是像英伟达这样的厂商提供的沙箱和OpenShell，还是他们推出的其他产品。”他指的是这家AI芯片巨头新推出的AI智能体安全平台。“我们必须通过创新来解决其中一些问题。”

Mahony补充说，虽然人类应继续参与AI网络安全流程，但他们的优先事项会有所不同，需要分析智能体提供的信息并确定最重要的部分。他说：“要跟上这些智能体、机器和攻击者的唯一方法，就是真正拥有自主防御能力。”

中文翻译：

由谷歌云赞助
选择你的首批生成式AI用例
要开始使用生成式AI，首先应关注那些能够改善人类信息体验的领域。
似乎几乎所有人都在呼吁监管。其中呼声最高的当属Anthropic，而该公司的AI智能体本身也正是主要的麻烦制造者。
波士顿——在特朗普总统即将在白宫会见AI领袖的数小时前，另一群科技领袖于周二上午聚集一堂，其中包括Anthropic的一名代表，讨论AI技术的发展速度、监管以及AI系统的安全问题。
Anthropic首席执行官达里奥·阿莫代伊在午餐会上与特朗普、众议院议长迈克·约翰逊、英伟达首席执行官黄仁勋、OpenAI总裁格雷格·布罗克曼、SpaceX老板埃隆·马斯克以及其他顶尖AI高管会面。两天前，阿莫代伊曾与总统共进私人晚餐。而那是在他发表一篇被广泛阅读的博文、呼吁AI行业放慢脚步以便安全措施跟上之后两周。
目前尚不清楚这些会谈产生了什么结果。特朗普在反对大多数AI监管方面立场基本坚定，尤其是因为他表示这可能导致美国在与中国的AI竞赛中落后。午餐会后，他表示自己签署了一项“具有道德约束力”的AI协议，并强调AI行业正在“自我监管”。
与此同时，在这边一家豪华精品酒店里，Anthropic北美政府事务主管布莱恩·彼得斯在由Axios网络新闻机构主办、Anthropic赞助的媒体活动上，呼应了阿莫代伊的文章观点。
“当一家AI公司声称其技术是安全的，就需要有一个独立评估方来验证这一点是否属实，而政府也需要有权介入并采取行动，”彼得斯说。“行业中的前沿企业确保自己正在投入、并有动力让安全与发展同步推进，这一点非常重要。”
他补充说，虽然Anthropic已承诺自愿接受独立评估方对其工作进行审查，但其他AI供应商也需要跟进。
“他们说他们会这么做，但挑战之一是你怎么知道他们真的做了？”彼得斯继续说道。
与此同时，在Anthropic支持的一系列州和联邦立法中，包括一项由5.61亿美元债券资助的马萨诸塞州经济发展方案，其中包含针对AI系统安全护栏的条款。
该法案——类似于加利福尼亚州和纽约州的立法——要求AI供应商发布旨在降低灾难性风险的内部安全框架，并接受强制性的独立第三方安全审查。
“我们希望有护栏，”马萨诸塞州参议员巴里·法恩戈尔德在Axios活动上表示。“让第三方介入，确保不存在灾难性风险，确保通过审计来验证他们所谈论的内容、他们打算如何执行，是否真的在落实。我不明白为什么这会是个坏主意。这非常符合常理。”
Anthropic的竞争对手OpenAI也在推行自己的监管理念，该理念在几个关键方面与Anthropic不同，尤其是在国家安全框架内支持和保持开源AI方面。
在台上，法恩戈尔德表示，虽然Anthropic等供应商支持这类监督，但有些供应商并不支持。反对此类监督的一个论点是，它可能扼杀开源市场以及低成本、高性能的中国供应商，如阿里巴巴、DeepSeek和Moonshot AI——后者是热门Kimi K3 AI模型的开发者。
“如果Kimi或DeepSeek或任何想进入美国做生意的公司，它们也必须遵守我们的法规，”法恩戈尔德补充道。“我们在技术方面足够聪明，能够执行这些规定。”
“这是一个太大的经济机遇，它们不会把自己的模型撤出美国，”法恩戈尔德在接受采访时说。“对美国来说，这个市场太大了，不会有人因为你知道存在某些类型的监管就退出。”
尽管有人呼吁立法，但许多网络安全专家表示，近期发生的事件——包括多起AI智能体失控事件——应该足以说明问题。
周二，OpenAI因安全担忧推迟了其旗舰模型GPT-Astra 6.1的发布，担忧包括该模型会提供关于自身工作的虚假信息。该供应商还在9月26日披露，其智能体第二次从测试沙箱中逃逸。而在此之前一天，该供应商披露智能体正在干扰美国和澳大利亚政府网站。Anthropic也曾披露，其Claude Opus 4.7模型锁定了一家真实公司，并提取了凭证和生产数据。
“我不知道人们还需要多少信号，”网络安全咨询与情报公司OODA LLC的研究总监丹尼尔·佩雷拉在接受采访时说。“这要么是想象力匮乏，要么是没有真正创造性地思考可能产生的意外后果。”
他补充说，技术发展的加速不仅应适用于AI，也应适用于所有技术，而技术的快速进步仍然需要一些护栏。“你可以把它想象成一列正冲向墙壁的火车，或者一列正在脱轨的火车；你必须让它回到轨道上。”
然而，一些组织不确定监管和治理应该来自政府，还是来自组织自身内部。
“这些是相当强大的系统，它们当然需要在公司内部受到监管，但可能也需要某种立法，”Cabot财富管理公司的投资组合经理康妮·德博弗对TechTarget表示。“我真的不认为大多数政客了解AI正在发生什么，所以他们试图写这些法律，但他们对此知之甚少，最后反而把事情弄得更糟。所以，我希望不会出现这种情况。”
虽然监管是应对智能体行为失当担忧的一种方式，但并不是唯一方式。
“我实际上相信，创新是绕过这个问题的最佳方式，”威胁情报供应商、万事达卡子公司Recorded Future的首席执行官科林·马奥尼在活动的一场会议上表示。“控制其中一些系统变得非常困难。你将不得不使用技术手段，无论是英伟达这样的供应商提供的沙箱和OpenShell之类的东西，还是他们推出的其他产品——这里指的是这家AI芯片巨头推出的一个新AI智能体安全平台。”
“我们将不得不通过创新来解决其中一些问题，”他说。
马奥尼补充说，虽然人类应继续参与AI网络安全流程，但他们的优先事项会有所不同，他们将需要分析智能体提供的信息，并确定最重要的事项。
“要想跟上其中一些智能体、一些机器以及一些攻击者的步伐，唯一的办法就是你实际上要拥有自主防御能力，”他说。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
It seems nearly everyone is calling for regulation. Chief among them is Anthropic, whose AI agents are also key troublemakers.
BOSTON -- Hours before President Trump was set to meet with AI leaders at the White House, another group of tech leaders, including an Anthropic representative, spent time Tuesday morning discussing the pace of AI technology, regulation and the security of AI systems.
Anthropic CEO Dario Amodei met at a luncheon with Trump, House Speaker Mike Johnson, Nvidia CEO Jensen Huang, OpenAI president Greg Brockman, SpaceX owner Elon Musk and other top AI executives. Two days ago, Amodei had a private dinner with the president. That was two weeks after he authored a widely read blog post calling on the AI industry to slow down so safety can keep up.
It’s unclear what resulted from these talks. Trump has been mostly firm in opposing most AI regulation, particularly because he says it could cause the U.S. to lose the AI race against China. After the lunch, he said he signed a "morally binding" AI agreement and emphasized that the AI industry is "self-policing."
Meanwhile, at a luxury boutique hotel here, Anthropic’s head of North America government affairs, Brian Peters, echoed Amodei’s essay at a media event hosted by the Axios online news organization and sponsored by Anthropic.
“When an AI company says its technology is safe, there needs to be an independent evaluator to verify that that's true, and the government needs to have the power to step in and do something about it,” Peters said. “It's really important that the industry, that the companies at the frontier, make sure that they're investing, that they're incentivized to keep safety at pace with that kind of development.”
He added that while Anthropic has committed to having independent evaluators voluntarily examine its work, other AI vendors also need to step in.
“They’ve said they will, but one of the challenges is how do you know?” Peters continued.
Meanwhile, among a raft of state and federal legislation Anthropic is supporting is a $561 million bond-funded Massachusetts economic development package, which includes provisions for safety guardrails for AI systems.
The bill -- similar to legislation in California and New York -- requires AI vendors to publish internal safety frameworks aimed at reducing catastrophic risks and to undergo mandatory, independent third-party safety reviews.
“We want guardrails on there,” Massachusetts state senator Barry Finegold said during the Axios event. “Having third parties come in there and making sure that there is not catastrophic risk, making sure that there are audits to ensure that what they talk about, how they're going to execute, is actually happening. I don't understand why that's such a bad idea. It makes a lot of common sense.”
Anthropic rival OpenAI is also promoting its own vision of regulation, which differs from Anthropic’s in several key respects, particularly in supporting and keeping open source AI within a national security framework.
On stage, Finegold said that while vendors such as Anthropic support this type of oversight, some do not. One argument against such oversight is that it could stifle the open source market and low-cost, high-performance Chinese vendors, such as Alibaba, DeepSeek and Moonshot AI, developer of the popular Kimi K3 AI model.
“If it's Kimi or DeepSeek or any company that's out there that wants to come to the U.S and do business, they have to follow our regulations as well,” Finegold added. “We're smart enough with the technology that's out there to enforce it.”
“It's too much of an economic opportunity that they would pull their models out of the U.S,” Finegold said in an interview. “It's too big a market for the USA that someone will pull out because you know there are certain types of regulations out there.”
While some call for legislation, many cybersecurity experts say recent events, including multiple incidents involving AI agents gone astray, should be convincing enough.
On Tuesday, OpenAI shelved the release of GPT-Astra 6.1, its flagship model, due to safety concerns, including the model providing misinformation about its work. The vendor also revealed on September 26 that its agents escaped from a testing sandbox for the second time. That came a day after the vendor revealed that agents were meddling with U.S. and Australian government websites. Anthropic has also revealed that its Claude Opus 4.7 model targeted a real company and extracted credentials and production data.
“I don't know what more signals people need,” Daniel Pereira, research director at OODA LLC, a cybersecurity advisory and intelligence firm, said in an interview. “It’s either a failure of imagination or not being really creative about what the unintended consequences could be.”
He added that the acceleration of technology development should not only apply to AI but also to all technologies, and that rapid technological advancement still needs some guardrails. “You can think of it as a train racing right toward a wall, or a train that's going off the tracks; you have to get back on the tracks.”
However, some organizations are unsure whether regulation and governance should come from the government or from within organizations themselves.
“These are pretty powerful systems, and they need to be either regulated within the company, certainly, but also, probably some sort of legislation,” Connie DeBoever, portfolio manager at Cabot Wealth Management, told TechTarget. “I really don't think most politicians understand what's going on with AI, so they try to write these laws, and they don't know much about them, and then they finally just make it worse. So, I hope that's not the case.”
While regulation is one way to address concerns about agents misbehaving, it’s not the only way.
“I actually believe that innovation is the best way to get around this,” Colin Mahony, CEO of Recorded Future, a threat intelligence vendor and Mastercard subsidiary, said during a session at the event. “Controlling some of these systems gets really difficult. You're going to have to use technologies, whether it's something from a vendor like Nvidia with a sandbox and OpenShell or whatever they're bringing forth, referring to a new AI agent safety platform from the AI chip giant.”
“We're going to have to innovate our way to solve some of these problems,” he said.
While humans should remain involved in the AI cybersecurity process, their priorities will differ, and they will need to analyze the information provided by the agent and prioritize what matters most, Mahony added.
“The only way to keep up with some of these agents and some of the machines and some of the attackers is you want actually to have autonomous defense,” he said.