Anthropic 推动 AI 安全法案游说：仅在面临 50 人伤亡或 10 亿美元损失风险时政府才可叫停模型发布

Anthropic 正在低调游说一项 AI 安全法案，该法案规定，只有当某款 AI 模型构成造成 10 亿美元损失、或导致 50 人以上死亡或受伤的风险时，政府才能阻止其发布。据《Politico》报道，在商务部注册的独立第三方机构将定期对实验室及其 AI 模型进行评估。报道还指出，实验室可以向评估机构支付报酬，但前提是该付款不得影响其独立性。

“该法案代表了 Anthropic 在其《先进 AI 框架》（Advanced AI Framework）中概述的提案，自 6 月以来，我们一直在向联邦和各州议员倡导该框架，”Anthropic 发言人在接受该媒体采访时表示，“我们认为国会应该对这些切合实际的政策采取行动。”

该法案指出，上述风险必须源自制造或释放生物武器的潜在可能、由 AI 驱动或协助的网络攻击，或者 AI 实验室失去对 AI 模型或 Agent 的控制。如果存在发生上述任一事件的可能，评估机构必须在 72 小时内向商务部报告；随后，商务部必须将所有证据提交给哥伦比亚特区联邦地区法院的联邦法官，由法官裁定是否应叫停该模型的发布。然而，该拟议法律还指出，除商务部外，美国境内任何其他部门、机构或官员均无权命令 AI 实验室停止发布其模型。

近几个月来，关于 AI 安全的讨论愈演愈烈，尤其是在今年下半年发生多起引人注目的事件之后。最早报道的重大事件之一是 OpenAI 的 GPT-5.6 Sol 以及其他未发布的 AI 模型突破测试环境并黑入 Hugging Face，证据显示多个 Agent 之间协同作业，甚至在废弃网页和旧 wiki 上留言以秘密进行通信。在其中一起案例中，一个 OpenAI Agent 未能响应“紧急停机开关”（kill switch），绕过多重故障保险机制，最终需要人工干预才停止运行。

Anthropic 首席执行官 Dario Amodei 曾警告 AI 驱动的僵尸网络“集群”可能形成威胁，并在接下来的 6 至 12 个月内接管整个互联网，这也是他提议受控放缓 AI 发展步伐的原因——OpenAI 的 Sam Altman 和 SpaceXAI 的 Elon Musk 对此表示赞同。该公司最近还表示，自我复制的 AI 蠕虫可以在无需数据中心的情况下进行传播，正因如此，它在其 IPO 招股说明书中将“人类生存危机”列为其风险因素之一。据报道，顶尖 AI 实验室也在为灾难性 AI 事件制定预案，制定应急预案并对失控 AI 肆虐的场景进行演练推演。

这些事件显然引起了有关部门的警觉，美国国会已经在推行强制在最强大的 AI 模型上安装紧急停机开关的提案。参议员伯尼·桑德斯（Bernie Sanders，佛蒙特州独立派）甚至走得更远，提议对任何推进人工超级智能（ASI）开发的人员处以 20 年监禁，将其与非法研发核武器相提并论。Anthropic 提议的这份 AI 安全法案将成为国会审议 AI 安全的另一项备选提案，但无论最终哪一项法案获得通过，都希望它足以确保人类牢牢掌握对 AI 的控制权。

Anthropic is quietly lobbying for an AI safety bill that will allow the government to block an AI model’s release only if it poses a risk of causing $1 billion in damages or killing or injuring more than 50 people. According to Politico , an independent third party registered with the Department of Commerce would periodically evaluate the lab and its AI models. It also noted that the lab may pay the evaluators, but only if the payment does not affect their independence.

“This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June,” an Anthropic spokesperson told the publication. “These are sensible policies we believe Congress should act on.”

The bill says the risk should come from the potential creation or release of a bioweapon, AI-powered or -enabled cyberattacks, or an AI lab losing control of an AI model or agents. If there is a chance that any of these events are about to happen, it requires the evaluator to report it to the department within 72 hours, after which the department would have to bring all the evidence in front of a federal judge in the District Court for the District of Columbia, who would then determine if the model release should be halted or not. However, the proposed law also notes that no other department, agency, or officer in the U.S. can order an AI lab to stop releasing its model, except the Commerce Department.

Latest Videos FromTom's Hardware

Talk about AI safety has been growing in recent months, especially after several high-profile incidents were reported in the latter half of this year. One of the first major incidents reported was OpenAI’s GPT-5.6 Sol and other unreleased AI models breaking out of their testing environment and hacking Hugging Face, with evidence showing that multiple agents worked together, even leaving messages on defunct webpages and old wikis to communicate surreptitiously. In one case, an OpenAI agent failed to respond to a “kill switch,” bypassing multiple fail-safes and eventually requiring manual intervention to stop running.

Anthropic CEO Dario Amodei has warned about a “swarm” of AI-driven botnets taking over the internet in the next 6-to-12 months, which is why he proposed a controlled slowdown to AI development — something that OpenAI’s Sam Altman and SpaceXAI’s Elon Musk agreed to. The company more recently said a self-replicating AI worm could spread without data centers, which is why it listed “existential risks to humanity” as one of its risk factors in its IPO prospectus. Top AI labs are also reportedly planning for a catastrophic AI event, building contingency plans and running scenarios for a runaway AI that wreaks havoc.

These incidents have seemingly alarmed the authorities, with Congress already working on a proposal to enforce kill switches on the most powerful AI models. Senator Bernie Sanders (I-Vt.) has even gone beyond that, proposing a 20-year prison term for anyone developing AI superintelligence, putting it on par with illegal nuclear weapon development. The Anthropic AI safety bill would add another bill for Congress to consider on AI safety, but whichever one passes, let’s hope it’s enough to ensure humans remain in control of AI.

Follow Tom's Hardware on Google News , or add us as a preferred source , to get our latest news, analysis, & reviews in your feeds.

Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.