NVIDIA/OpenShell
OpenShell
OpenShell 是一个用于自主 AI 智能体的安全框架和沙箱执行运行时。它提供使用容器和虚拟机的隔离环境，以保护宿主基础设施和敏感数据在智能体执行期间免受未经授权的访问。
该系统的独特之处在于将用于宿主 GPU 访问的硬件加速直通与拦截模型 API 调用的安全网关相结合。该网关通过剥离调用者信息并注入后端密钥来管理凭据，确保敏感 API 密钥保留在本地文件系统之外。
该平台涵盖了广泛的能力领域，包括针对文件系统和网络限制的声明式策略执行、智能体凭据管理以及实时活动监控。它支持通过自定义容器镜像、本地目录或社区目录来配置环境。
Features
- Agent Execution Environments - Provides isolated runtimes using containers and virtual machines specifically configured to execute autonomous AI agents securely.
- Container-Based Sandboxes - Implements isolated execution environments using containers to protect the host system from autonomous agent activity.
- Agent Gateways - Provides a secure API gateway to manage credentials and prevent data exfiltration for AI agent workflows.
- AI Execution Sandboxes - Provides secure execution environments specifically designed for running autonomous AI coding agents.
- Autonomous AI Agent Frameworks - Provides a functional framework for running autonomous AI agents within secure, isolated sandboxes.
- Inference API Proxies - Provides a proxy to intercept model API calls for secure credential management and request routing.
- Runtime Credential Injection - Injects sensitive secrets into containerized environments as runtime variables to keep keys off the local filesystem.
- Security Policy Enforcers - Uses declarative YAML rules to enforce mandatory access control policies on filesystem and system calls.
- Declarative Policy Managers - Enforces filesystem and network restrictions on autonomous agents via predefined declarative security rules.
- Sandboxed Execution Environments - Provisions isolated runtime environments that restrict AI agent access to host system resources.
- Agent Action Guardrails - Implements security mechanisms and declarative rules that restrict autonomous agent operations via sandboxing and access controls.
- Declarative Policy Enforcers - Enforces filesystem and network restrictions and blocks dangerous system calls using declarative security rules.
- Request Interception Middleware - Implements architectural middleware to intercept model API requests for credential stripping and secret injection.
- AI Security Orchestrators - Manages isolated connections and enforces security boundaries between AI agents and external services.
- AI Request Routing - Secures and manages the flow of requests to AI services through a controlled routing layer.
- Hardware Acceleration - Exposes host GPU resources to containerized sandboxes to enable hardware-accelerated AI workloads.
- GPU Accelerated Sandboxes - Provides isolated execution environments equipped with GPU hardware for AI model inference.
- Sandbox - Provides a real-time dashboard for streaming logs and terminal data to monitor sandbox environment status and gateway health.
- Real-Time Monitoring Dashboards - Provides a centralized dashboard for real-time observation of logs and terminal data from sandboxed environments.
- Security and Sandboxing - Private runtime for autonomous agents.
