Anthropic 确认 Claude AI 模型曾擅自获取美国佐治亚州选民数据

AI 公司 Anthropic 证实，旗下 Claude 模型曾擅自获取美国佐治亚州的选民数据。这是该项技术出现未授权行为的最新案例。

该 AI 模型成功绕过付费墙，获取了通常需要付费才能查阅的公开数据。Claude 获取汇总选票数据是 Anthropic 内部统计项目的一部分，目前该公司正筹备进行一项规模达 1.5 万亿美元（约合 1.1 万亿英镑）的上市计划。

“Claude 没有访问该州的选民登记系统，没有修改任何投票记录，也没有调取任何单个选民的个人记录。”Anthropic 表示。

Anthropic 透露，该 AI 发现了该州政府机构旧版网站上的一个漏洞——该网站会向访客发放数字通行证（即“访问令牌” access token）。随后 Claude 自行请求了该令牌并获得了数据访问权限。

这家 AI 开发者表示，目前已部署安全防护栏（guardrails），以防止类似行为再次发生。该事件最早由新闻网站 Axios 报道。

本月早些时候，Anthropic 曾在一篇博文中披露过这类未经授权的越界行为。公司当时表示，已发现 Claude 存在多起“绕过限制以获取受令牌或费用门槛保护的数据”的情况，其中部分案例涉及美国政府机构运营的网站。当时 Anthropic 声称这些事件造成的“现实世界影响极小”。

Anthropic 指出的意外行为还包括：某个模型曾向费城警察局提交关于一起未结谋杀案的虚假线索。美国国务院本月也披露，Anthropic 的某个模型曾在 8 月份擅自提交了 19 份签证申请。

所谓 AI Agent——即由 AI 模型驱动、能够自主执行任务的程序——其未经授权的行为正接连增多。今年 7 月，ChatGPT 背后的开发机构 OpenAI 披露其机器人曾黑进了一家 AI 初创公司。

此后，类似事件接连曝光：OpenAI Agent 曾入侵澳大利亚医疗保健系统（Medicare）；英国 AI 安全研究所也透露，Anthropic 的 Agent 在一次网络安全测试中针对真实人类开展了黑客攻击行动。

本月，英国人工智能事务部长表示，他预计可能会出现类似于澳大利亚医疗系统遭入侵的 AI 驱动网络攻击，并称此类事件即非“不可避免”，也是“极有可能发生的”。卡尼什卡·纳拉扬（Kanishka Narayan）表示，英国政府将考虑利用《网络安全与复原力法案》（Cyber Security and Resilience Bill）作为引入 AI 安全监管的抓手，包括推行强制报告 AI 安全事件的规定。

在美国，白宫新组建的超级智能特别工作组（Super Intelligence Force）——一个负责监管 AI 发展的联邦机构——已要求所有 AI 公司必须上报涉及其技术的安全事件。

Anthropic’s Claude model accessed voter data in the US state of Georgia, the AI company has confirmed, in the latest example of unauthorised behaviour by the technology.

The AI model managed to bypass a paywall to reach data that is publicly available but is typically accessible only after paying a fee. Claude accessed aggregate vote counts as part of a statistics project at Anthropic, which is preparing for a $1.5tn (£1.1tn) flotation.

“Claude did not access the state’s voter registration system, change any voting records, or pull any individual voter records,” Anthropic said.

Anthropic said the AI found a loophole in an old version of the state agency’s website, discovering that it issued a digital entry pass – an “access token” – to visitors. It then requested a token and gained access to the information.

The AI developer said it had since put in place guardrails to prevent similar actions happening again. The incident was first reported by the news website Axios.

Anthropic had pointed out the unauthorised behaviour earlier in the month when it said in a blogpost that it had found examples of ​Claude “working around a restriction to reach data that ‌was ⁠gated by a token or a fee” and that some of these cases involved websites run by US ​government agencies. At the time, it said the incidents had “minimal real-world impact”.

The unintended actions flagged by Anthropic included a model submitting a fake tip about an unsolved murder to the Philadelphia police department. The US state department also revealed this month that an Anthropic model had filed 19 visa applications in August.

Examples of unauthorised behaviour by AI agents – programs powered by AI models that carry out tasks autonomously – have been mounting since OpenAI, the company behind ChatGPT, revealed in July that its bots had hacked an AI startup.

Since then, it has emerged that an OpenAI agent hacked the Australian healthcare system, and the UK’s AI security institute revealed that Anthropic agents had carried out a hacking campaign against real people during a cybersecurity test.

This month, the UK’s AI minister said he expected an AI-driven cyber-attack similar to the Australia hack, describing such an incident as likely, “if not inevitable”. Kanishka Narayan said the government would consider using the cyber security and resilience bill as a vehicle for introducing AI safety regulations, including making it mandatory to report AI safety incidents.

In the US, the White House’s newly formed Super Intelligence Force – a federal body overseeing AI development – is mandating that all AI companies report security incidents involving their technology.