Unverified60% confidenceEventExact time
In 2023, researchers demonstrated hiding malicious instructions in invisible text areas of web pages, successfully hijacking a GPT-4-based browser agent into performing unauthorized operations.
2
Sources
60%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Related Claims
UnverifiedMultiple public cases in 2024 have demonstrated the feasibility of prompt injection attack vectors, including hijacking browser-controlling Agents via malicious web pages to execute fund transfers74% similarUnverified2023年安全研究员Riley Goodside与Johann Rehberger等人演示了通过网页内容、邮件正文劫持GPT-4插件和AutoGPT的完整攻击链72% similarUnverified2023年，安全研究者Johann Rehberger展示了如何通过在网页中嵌入不可见文本，让ChatGPT在用户不知情的情况下将对话内容泄露给外部服务器68% similarUnverifiedAllowing third-party software of unknown origin to bridge web models for reading and writing local files and controlling the mouse and keyboard poses security risks, and may steal screen content, clipboard data, and even local files.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/602677API
curl https://kongchang.com/api/v1/knowledge/claims/602677MCP
get_claim(id=602677)