How autonomous AI agents operate inside enterprise environments without ever touching unauthorized systems
OpenShell Sandbox — Hardware-Level Isolation
1
用户请求接收
Agent receives task via NemoClaw CLI or API. Request logged and threat-scanned before any execution begins.
Ingress Gate
↓
2
能力下降应用
Filesystem lock, hot-reloadable network policy, and capability restrictions activated at kernel level. Every access path not explicitly whitelisted is hardware-blocked.
Hardened
↓
3
仅授权资源
Agent accesses only pre-approved internal directories, databases, and API endpoints. No lateral movement possible — each permission is explicitly granted per session.
Scoped Access
↓
4
实时数据包过滤
All outbound network traffic inspected in real time. Anomalous packets flagged and dropped. On violation detection: agent kernel triggers immediate self-healing shutdown.
Edge Filter
↓
5
受保护的执行和输出
Task completes inside sandbox. Results returned to authorized user. Zero IP exfiltration possible — the agent never has the ability to reach outside its permitted scope.
Secure Output
🚫 永久阻止
✕Access to unauthorized filesystem directories
✕Outbound connections to non-whitelisted IPs/ports
✕Privilege escalation or kernel manipulation
✕Exfiltration of source code, financial data, IP
✕Prompt injection attacks from external content
✓ 允许的操作
✓Read/write to explicitly authorized directories
✓API calls to whitelisted internal services only
✓Session-scoped memory — cleared after each task
✓Audit logging of all actions for compliance review
✓Full output to authorized user only
企业用例 — NemoClaw 安全解锁
🤖 Autonomous coding assistant
📊 Financial transaction analysis agent
🔍 Legal document review & summarization
📧 Email drafting with internal data access
🏭 Manufacturing anomaly detection
🔐 Security audit automation
📋 Compliance report generation
Source: NemoClaw & OpenShell Sandbox Hardening User Guide — NVIDIA Product Security Team