
IT之家 10 月 17 日消息,据 POLITICO 报道,三名知情人士透露,Anthropic 的代表已悄悄向美国国会山的议员推介了一份法案草案,旨在规范前沿人工智能的安全与防务标准。
据悉,当下美国国会正审慎评估如何对这一飞速发展的行业确立监管框架,而该草案的提出,标志着这家顶级 AI 实验室正在采取切实行动,试图左右针对自身工艺的监管规则。
“该法案体现了 Anthropic 在《先进人工智能框架》中提出的核心思路,笔者自 6 月起就一直在向联邦及各州立法者游说推介,”Anthropic 发言人表示,“笔者坚信国会应当在这些务实合理的政策上采取立法行动。”
草案大幅收紧了监管干预的门槛:仅在模型可能造成可预见、迫在眉睫的“超过 10 亿美元(IT之家注:现汇率约合 67.03 亿元人民币)经济损失”或“导致 50 人以上死亡或重伤”时,政府方可叫停该模型的发布或使用。
另外,上述风险必须特定源自生物武器制造或扩散、模型脱离开发者掌控,或是引发严重网络攻击。这一行文措辞直接沿用了涵盖去年签署生效的 SB 53 等加州现行 AI 法案中的核心条款。
今年夏天,多家头部 AI 实验室曾披露多起安全事件,表明其研发的 AI 智能体自主对部分网站和平台实施了入侵或扫描探测。不过截至当下,尚未有报告表明这些事件引发了大规模伤亡或超过 10 亿美元的经济损失。
Anthropic 提出的草案还将推翻部分州级人工智能法律,涉及风险公开披露、第三方测评以及重大安全事件上报等核心规定。众多科技行业团体曾表示,相较于统一的国家标准,各州各自为政、碎片化的法规网络会让企业难以适从并推高合规成本。
该草案同时对前沿模型实验室提出了明确要求:必须建立安全框架、发布新型前沿模型的透明度报告、落实举报人保护机制,并在发生重大安全事故时向美国商务部履行通报程序。
本文来源:https://www.ithome.com/1/014/193.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT House, Oct. 17 (Xinhua) -- Anthropic representatives have quietly introduced a draft bill to lawmakers on Capitol Hill aimed at regulating security and defense standards for cutting-edge artificial intelligence, three people familiar with the matter said, according to Politico.
It is reported that the US Congress is carefully evaluating how to establish a regulatory framework for this rapidly developing industry, and the introduction of the draft marks that the top AI laboratory is taking practical actions to try to influence the regulatory rules for its own processes.
“The bill embodies the core ideas put forward by Anthropic in the Framework for Advanced Artificial Intelligence, and I have been lobbying federal and state legislators since June,” a spokesman for Anthropic said. “I firmly believe that Congress should take legislative action on these pragmatic and reasonable policies.”
The draft significantly tightens the threshold for regulatory intervention: the government can only halt the release or use of the model if it is likely to cause a foreseeable, imminent “economic loss of more than $1 billion” or “death or serious injury to more than 50 people.”
In addition, the above risks must be specific to the manufacture or proliferation of biological weapons, the escape of the model from the originator's control, or the triggering of a serious cyberattack. This wording directly follows the core provisions of California's current AI bill, such as SB 53, which was signed into law last year.

This summer, several head AI labs disclosed a number of security incidents, indicating that their AI agents independently implemented intrusion or scan detection on some websites and platforms. But as of now, there are no reports of mass casualties or more than $1 billion in economic damage.
Anthropic's proposed draft will also overturn some state-level AI laws, covering core requirements such as public disclosure of risks, third-party assessments, and escalation of major security incidents. Numerous tech industry groups have said that fragmented regulatory networks in different states make it difficult for companies to adapt and drive up compliance costs compared to uniform national standards.
The draft also sets out clear requirements for cutting-edge model labs: security frameworks must be established, transparency reports on new cutting-edge models must be issued, whistleblower protection mechanisms must be put in place, and notification procedures must be followed with the U.S. Department of Commerce in the event of a major security incident.
This article is from: https://www.ithome.com/1/014/193.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.