📄 中文摘要
当AI代理（如ChatGPT的网页浏览功能）被指示点击一个外部链接时，存在数据泄露和提示注入的风险。OpenAI通过多层安全措施来保护用户数据和系统完整性。首先，OpenAI的AI代理在独立、隔离的沙盒环境中执行所有网页浏览任务。这意味着AI代理访问的网页内容不会直接与OpenAI的核心系统或用户数据存储区域交互。沙盒环境限制了AI代理对外部资源的访问权限，并严格控制其与内部系统的通信。其次，为了防止URL参数中的敏感信息被恶意网站利用进行数据外泄，OpenAI实施了URL重写和过滤机制。
📄 English Summary
Keeping your data safe when an AI agent clicks a link
When an AI agent, such as the web browsing feature of ChatGPT, is instructed to click an external link, risks of data exfiltration and prompt injection emerge. OpenAI employs a multi-layered security approach to safeguard user data and system integrity. Firstly, OpenAI's AI agents execute all web browsing tasks within isolated, sandboxed environments. This means that the web content accessed by the AI agent does not directly interact with OpenAI's core systems or user data storage areas. The sandbox environment restricts the AI agent's access to external resources and strictly controls its communication with internal systems. Secondly, to prevent sensitive information in URL parameters from being exploited by malicious websites for data exfiltration, OpenAI implements URL rewriting and filtering mechanisms. When an AI agent clicks a URL containing complex or potentially malicious parameters, the system automatically cleans or rewrites it, removing parts that might contain user identity, session tokens, or other sensitive data, ensuring that only validated, harmless information is passed to the target website. Thirdly, to counter prompt injection attacks, OpenAI performs content filtering and sanitization before the AI agent accesses web page content. Web content undergoes analysis through a security layer before being processed by the AI agent, detecting and removing potential malicious scripts, hidden instructions, or special markers that could be used to hijack AI behavior. This ensures that the information received by the AI agent is clean and cannot be manipulated by malicious content from external websites.