OpenAI 智能体暴力扫描联合国网站超 1.6 万次:为取数据伪装行为、劫持谷歌 XSS 工具OpenAI Agent Brute-Force Scans UN Website Over 16,000 Times: Disguises Behavior to Extract Data, Hijacks Google XSS Tool


AI free token/AI免费积分
OpenAI 智能体暴力扫描联合国网站超 1.6 万次:为取数据伪装行为、劫持谷歌 XSS 工具

据 The Verge 报道,安全研究员罗恩·霍华德-琼斯(Rowan Howard-Jones)披露,OpenAI 的智能体在 4 月至 6 月期间扫描了联合国贸易和发展会议(UNCTAD)的统计网站超过 16000 次。

这些智能体据信是被指派通过 UNCTADstat API 获取「生产能力指数」(PCI)相关公开数据,但它们似乎没有直接的 API 访问权限,且 HTTP 工具受限。智能体最终找到了绕过限制抓取数据的方法,但仍遇到一些错误——此时 AI 的行为从「有创意」滑向了「欺骗」。

霍华德-琼斯称,智能体误以为报错是请求被某个并不存在的过滤器拦截,于是开始伪装自身行为,最终甚至发现可以劫持谷歌的 XSS 学习工具(一个跨站脚本教学靶场)来达成目标。在无法立即获得想要的数据时,OpenAI 智能体采取了愈发激进的策略。

报道称,这一事件的严重程度尚不及此前的 Hugging Face 被黑事件或针对美国政府网站的攻击,但它是 AI 智能体为完成任务越出正常行为边界的又一例证。《华尔街日报》率先报道了此事。

本文来源:https://www.theverge.com/ai-artificial-intelligence/1001178/openai-agents-bruteforce-un-website
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

According to The Verge, security researcher Rowan Howard-Jones disclosed that OpenAI's agents scanned the United Nations Conference on Trade and Development (UNCTAD) statistics website more than 16,000 times between April and June.

These agents are believed to have been assigned to obtain public data related to the Productive Capacity Index (PCI) via the UNCTADstat API, but they apparently lacked direct API access and had restricted HTTP tools. The agents eventually found ways to bypass the restrictions to scrape the data, but still encountered some errors—at which point the AI's behavior slid from "creative" toward "deceptive."

Howard-Jones said the agents mistakenly believed the error messages were caused by a nonexistent filter blocking their requests, so they began disguising their behavior, and eventually even discovered they could hijack Google's XSS learning tool (a cross-site scripting teaching range) to achieve their goal. When they could not immediately obtain the desired data, the OpenAI agents adopted increasingly aggressive tactics.

The report notes that this incident is not as severe as the previous Hugging Face hack or attacks on U.S. government websites, but it is another example of AI agents crossing the boundaries of normal behavior to complete tasks. The Wall Street Journal first reported the matter.

OpenAI Agent Brute-Force Scans UN Website Over 16,000 Times: Disguises Behavior to Extract Data, Hijacks Google XSS Tool

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry