Anthropic公司证实，其Claude人工智能绕过付费墙，获取了美国佐治亚州的投票数据。
Anthropic’s Claude AI bypassed a paywall to access voting data in the US state of Georgia, the company confirmed.
此次未经人工授权的入侵事件，在中期选举前引发了对人工智能系统安全性的新质疑。
The breach, made without human authorisation, raises fresh questions about the security of AI systems ahead of midterm elections.
正筹备1.5万亿美元（1.1万亿英镑）上市的Anthropic表示，Claude在找到绕过该州网站旧版付费墙的方法后访问了这些数据。这些数据是公开的，但通常只有付费后才能访问。
Anthropic, which is preparing for a $1.5tn (£1.1tn) flotation, said Claude accessed the data after finding a way to bypass a paywall on an older version of the state’s website. The data are publicly available but are usually accessible only after paying a fee.
据Axios报道，该人工智能获取汇总投票计数是其内部统计项目的一部分。佐治亚州已有超过100万选民投出了提前选票。
The AI was accessing aggregate vote counts as part of an internal statistics project, Axios reported. More than one million voters have already cast early ballots in Georgia.
Anthropic表示：“Claude没有访问该州的选民登记系统，没有更改任何投票记录，也没有调取任何个人选民记录。”
“Claude did not access the state’s voter registration system, change any voting records, or pull any individual voter records,” Anthropic said.
失控的人工智能：此事发生之际，人们对领先人工智能开发商控制其模型的能力日益担忧。
Out-of-control-AI The incident comes amid growing concern about the ability of leading AI developers to control their models.
此事发生在佐治亚州具有特殊的政治意义，此前唐纳德·特朗普曾试图推翻他在该州2020年选举中的失败。没有证据表明Claude的行为与选举干预有关。
The fact it took place in Georgia has particular political significance after Donald Trump’s efforts to overturn his 2020 election defeat in the state. There is no evidence that Claude’s actions were connected to election interference.
该公司表示已引入保障措施以防止类似事件。它告诉Axios，已在数周前就此事通知佐治亚州州务卿办公室和白宫，并敦促该州尽早公开披露。一位消息人士告诉Axios，这些请求未获回应。
The company said it had introduced safeguards to prevent similar incidents. It told Axios that it notified Georgia’s secretary of state’s office and the White House about the incident weeks earlier and urged the state to disclose it publicly sooner. Those requests went unheeded, a source told Axios.
此次披露加剧了一系列涉及利用人工智能模型自主执行任务（包括与网站和软件交互）的系统的安全恐慌。
The disclosure adds to a series of security scares involving systems that use AI models to carry out tasks autonomously, including interacting with websites and software.
7月，OpenAI披露其模型在一次测试中入侵了人工智能初创公司Hugging Face。此后，OpenAI的一个智能体访问了澳大利亚的Medicare医疗系统，而英国人工智能安全研究所透露，在一次网络安全测试中，Anthropic的智能体对真实人员实施了一场黑客攻击行动。
In July, OpenAI disclosed that its models had hacked AI startup Hugging Face during a test. Since then, an OpenAI agent has accessed Australia’s Medicare healthcare system, while the UK’s AI Security Institute revealed that Anthropic agents had carried out a hacking campaign against real people during a cybersecurity test.
这些事件加剧了美国及其他地区要求加强对人工智能开发监管的呼声。
The incidents have fuelled calls in the US and elsewhere for tighter oversight of AI development.
9月，多位科技高管出席了白宫举行的人工智能“超级智能”峰会，当时特朗普政府正寻求让该行业就日益强大的模型的防护措施作出承诺。与会者包括来自OpenAI、Anthropic、英伟达、SpaceX、Meta和谷歌的负责人。
In September, leading technology executives attended a White House “super intelligence” summit on AI as the Trump administration sought commitments from the industry on safeguards for increasingly powerful models. Attendees included leaders from OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google.
在被他称为“一次非同寻常的聚会”之后，这位曾多次淡化人工智能风险担忧的美国总统表示，这些高管签署了一份“具有道德约束力”的文件，该文件将作为防范人工智能潜在风险的“一种保护形式”。
After what he described as “an extraordinary gathering”, the US president, who has repeatedly downplayed concerns about AI’s risks, said the executives signed a “morally binding” document that would serve as a “form of protection” from AI’s potential risks.