Plimsoll是一个自托管的代码沙箱工具，专为安全运行AI Agent生成的代码而设计1。该项目通过进程、容器、内核、虚拟机四层隔离级别来限制不可信代码的权限，用户可根据安全需求选择相应的隔离强度1。每次代码执行都会返回SHA-256哈希值作为执行完整性的加密验证1。
系统支持能力授予机制，允许AI生成的代码调用用户API，但凭证由代理端处理而不进入沙箱，从而避免敏感信息泄露1。Plimsoll已集成Trigger.dev TypeScript任务运行器，提供executeCode工具便于与AI框架配合1。项目支持多种提供商配置，包括基于QuickJS的WebAssembly环境、使用runsc或gVisor的Docker容器、e2b Firecracker虚拟机以及禁用模式1。此外，系统的会话模式可在多次调用间保持状态和变量1。
项目目前处于Pre-1.0阶段，由单一作者维护，尚未经过第三方安全审计1。
Plimsoll is a self-hosted code sandbox designed to safely execute code generated by AI agents.1 The tool implements four progressive isolation levels—process, container, kernel, and virtual machine—to restrict the permissions of untrusted code.1 Each code execution returns a SHA-256 hash serving as cryptographic proof of the run's integrity.1
The system enables AI-generated code to invoke user APIs without exposing credentials, as authentication is handled by a broker outside the sandbox rather than being passed into the isolated environment.1 Plimsoll currently integrates with Trigger.dev's TypeScript job runner and provides an executeCode tool for that framework.1 Users can choose from four execution providers: WebAssembly via QuickJS, Docker with runsc or gVisor containers, e2b Firecracker virtual machines, or a disabled mode.1 The platform also supports session-based execution that maintains state and variables across multiple invocations.1
The project remains in pre-1.0 stage and is maintained by a single author.1 It has not undergone third-party security auditing.1
评论
还没有评论，欢迎留下第一条。