Taming Agentic AI: Run Rogue Code Safely on Kubernetes with Agent Sandbox
summary
Kubernetes-native Agent Sandbox runs untrusted agent-generated code inside ephemeral pods backed by a CRD-based template system and a Python client SDK that calls a router/proxy to spin sandboxes. Default runtime is gVisor user-space kernel runtime (also supports Kata/Firecracker). It uses a warm pool (demo ran with two replicas) so execution starts in seconds, uploads code, runs it isolated, then deletes the pod — demo shows environment variables and CSV processing contained and Lovable runs it in production.
AI-written summary · about
chapters (6)
- 0:00Agent Sandbox project intro
- 1:50Why agents need sandboxing
- 3:55How Agent Sandbox works
- 5:58Live demo: code sandboxing
- 10:53Practical use cases
- 12:19Q&A: local dev & isolation
related talks
39mTrust, but Sandbox: Securing AI Agents in Kubernetes
29mSandbox for Agentic Application With Cloud Native Stack
09mBuilding an Agent Runtime With Open Infrastructure
25mPlease Sandbox your AI Agents
05mProject Lightning Talk: Run Sandbox Securely and Cost-Effectively With OpenKruise Agents
35mThe Enemy Within: Running Untrusted Code with gVisor
01h 42mHow to build and scale multi-agent AI systems
35mThe Enemy Within: Running Untrusted Code with gVisor