📄 中文摘要
OS级沙箱通过在内核级别约束代理进程来划定边界，但仅靠内核隔离不足以防止恶意操作。内核无法区分合法的文件写入与恶意的后门写入，而当代理具有合法的网络访问权限时，内核网络隔离也无法提供保护。应用层防御在更高的语义层面上运作，能够理解命令结构、Unicode攻击、信任来源和凭证等。这些防御措施能够有效识别和阻止潜在的恶意行为，确保在沙箱内的安全性。
应用层防御：在沙箱内阻止数据外泄
出处: Application-Layer Defense: Stopping Exfiltration Inside the Sandbox
发布: 2026年3月10日
📄 English Summary
Application-Layer Defense: Stopping Exfiltration Inside the Sandbox
OS-level sandboxes establish boundaries by constraining agent processes at the kernel level, but kernel isolation alone is insufficient to prevent malicious actions. It cannot differentiate between legitimate file writes and malicious backdoor writes. Additionally, when agents have legitimate network access, kernel network isolation fails to provide protection. Application-layer defenses operate at a higher semantic level, understanding command structures, Unicode attacks, trust provenance, and credentials. These defenses can effectively identify and mitigate potential malicious activities, ensuring security within the sandbox.