开发者Antoniomika发布了名为“lend”的工具,旨在解决Claude Code、Codex等远程AI编码代理在虚拟机运行时无法安全使用本地笔记本电脑凭证的难题1。传统方法通常面临两难困境:要么将GitHub令牌等敏感凭证复制到虚拟机中导致泄露风险,要么使用缺乏审计功能且仅支持密钥的SSH agent forwarding1。“lend”作为位于笔记本端的代理,拦截来自虚拟机的请求,并在本地执行命令,强制要求通过Touch ID或策略审批来处理需要权限的操作,从而避免将敏感令牌直接复制到远程环境1。
该工具的核心安全机制采用默认拒绝策略,仅允许白名单中的可执行文件和SSH密钥指纹访问,同时支持“租约(Leases)”功能以临时授权多次相同操作1。为了优化工作流程,开发者介绍了配合使用的终端复用器herdr和移动端客户端Ovis:herdr用于管理多个并行运行的Agent会话,而Ovis则允许用户通过手机或Tailscale远程审批Agent提出的非凭证类请求(如文件编辑、URL获取)1。尽管“lend”提供了额外的安全层,但它目前仅支持macOS和Linux系统,且不支持交互式PTY程序1。此外,已批准的命令仍以用户身份运行,若用户批准了shell等危险命令,其保护意义将随之丧失1。
Developer Antonomika has released "lend," a tool designed to address the security risks associated with running remote AI coding agents, such as Claude Code and Codex, on virtual machines. The core problem it solves is the inability of these agents to securely access local laptop credentials without exposing sensitive data. Traditional methods either involve copying credentials like GitHub tokens directly into the VM, which creates a significant leakage risk, or using SSH agent forwarding, which lacks audit trails and supports only keys rather than broader authentication needs 1.
Lend operates as a proxy on the user's laptop that intercepts requests from the remote VM. Instead of executing commands in the insecure environment, it runs them locally and mandates approval via Touch ID or predefined policies, ensuring that sensitive tokens never leave the secure device. The system employs a default-deny strategy, permitting execution only for whitelisted binaries and specific SSH key fingerprints, while also supporting temporary leases for repeated operations 1. To manage this workflow, the developer utilizes herdr for handling multiple parallel agent sessions and Ovis, a mobile client accessed via Tailscale, to approve non-credential requests such as file edits or URL fetches remotely 1.
Despite its benefits, lend has notable limitations. It currently supports only macOS and Linux and does not handle interactive PTY programs. Furthermore, because approved commands still execute under the user's identity, granting permission to dangerous utilities like shell interpreters can negate the tool's protective advantages 1.
评论
还没有评论,欢迎留下第一条。