yanacuti1121 MXC (Microsoft eXecution Container): chạy code không tin tưởng — model output, plugins, tools — trong sandbox có policy granular. Cross-platform, multiple isolation backends.
Install
npm install @microsoft/mxc-sdk
Usage (TypeScript SDK)
import { MXCClient } from '@microsoft/mxc-sdk'
const client = new MXCClient()
const result = await client.exec({
command: ['python', 'untrusted_script.py'],
filesystem: {
readOnly: ['/data/input'],
readWrite: ['/tmp/output']
},
network: {
outbound: false // no network access
},
timeout: 30000
})
JSON Config Schema (v0.6.0-alpha)
{
"version": "0.6.0-alpha",
"command": ["node", "agent_generated_code.js"],
"filesystem": {
"readOnly": ["./context"],
"readWrite": ["./output"]
},
"network": {
"proxy": null,
"outbound": false
},
"timeout": 30
}
Isolation Backends
| Backend | Platform | Strength |
|---|---|---|
| ProcessContainer | Windows (default) | lightweight |
| Bubblewrap | Linux (default) | namespace isolation |
| Seatbelt | macOS (default) | sandbox profiles |
| Windows Sandbox | Windows | full VM-lite |
| LXC | Linux | container |
| MicroVM (NanVix) | Linux | full VM |
| Hyperlight | Cross | micro-VM |
Sandbox Lifecycle
provision → start → exec → stop → deprovision
Hỗ trợ long-running sessions, không chỉ one-shot.
Dùng khi nào
- Chạy model-generated code an toàn
- Sandbox plugin/tool calls của agent
- CI/CD: test untrusted PR code
- Security research: execute malware sample
Caveat
"No MXC profiles should be treated as security boundaries currently" Đang alpha — dùng cho dev/research, không dùng làm hard security boundary.
Source
https://github.com/microsoft/mxc · MIT · +64⭐ today