How to sandbox AI agents with OpenShell | AgentOps Unlocked
Can you sandbox AI agents before a prompt injection exposes your internal credentials? Learn how OpenShell isolates autonomous agents using declarative security policies. When autonomous agents read incoming documents and emails, hidden prompt injections can instruct them to expose environment variables and execute unauthorized curl commands. Join Red Hat Developer Advocate Legare Kerrison to review how to contain agentic systems using OpenShell, an open source project from NVIDIA. We examine the three core architectural components: the gateway control plane, the in-sandbox supervisor with a default deny network proxy, and the command-line interface. See a demonstration using the OpenCode coding agent as OpenShell blocks an attempted credential exfiltration attack and restricts unverified web requests. We also explain how running OpenShell with Kata Containers on Red Hat OpenShift provides lightweight virtual machine isolation with a dedicated kernel, preventing shared host kernel breakout risks. 00:00 Security risks of agentic systems like Hermes and OpenClaw 00:13 How prompt injection phishing attacks hijack AI agents 00:48 Sandboxing AI agents with NVIDIA OpenShell 01:10 OpenShell gateway and pluggable compute backends (Docker, Podman, Kubernetes) 01:26 OpenShell supervisor and default-deny security proxy 01:47 OpenShell CLI: managing agent policies and live monitoring 01:56 Preventing credential theft in OpenCode AI agent 02:30 Restricting unauthorized web browsing requests 02:51 Kernel isolation and sandbox escapes with Kata Containers 03:22 Running secure AI sandboxes on Red Hat OpenShift 03:30 Managing credentials and environment variables with OpenShell providers 03:41 OpenShell quickstart commands Explore open source AgentOps and Red Hat security tools: 📂 Access the AgentOps GitHub repository → https://red.ht/github-agent-ops 🔒 Explore Red Hat security solutions → https://www.redhat.com/en/topics/security 🤖 Learn about Red Hat OpenShift AI → https://www.redhat.com/en/technologies/cloud-computing/openshift/openshift-ai Watch the previous episode of AgentOps Unlocked to learn about moving an AI agent from a laptop to a Kubernetes cluster → https://youtu.be/6MuAOFfJk2w #AgentOps #AIAgents #OpenShell #AISecurity #KataContainers #OpenShift #Kubernetes #OpenSource