Agentic AI Security#
AI agents combine model-generated decisions with tools that can read data, write files and interact with services. Securing an agent deployment therefore requires controls across inputs, tool execution, data handling and network access, not just the inference endpoint.
Agent Harness Installer starts with agent sandboxing. This documentation will grow as additional security capabilities are implemented.
Available security features#
Feature |
Scope |
Status |
|---|---|---|
Run OpenClaw, Hermes or Deep Agents Code in an OpenShell sandbox with restricted filesystem and network access |
Available; OpenShell 0.1.5 on Linux with Docker |
Start with the Agent Sandboxing getting started guide to install an agent, connect it to inference and verify its sandbox. Sandboxing is opt-in through runtime selection; host deployment remains the default. The installer supplies initial sandbox policies and inference configuration so users do not need to author a policy to get started.
Future areas#
Potential future areas include prompt protection and PII masking. They are not implemented security features of this integration today, and no delivery schedule is implied.
Sandboxing, prompt protection and PII masking address different security concerns and are intended to complement one another.