Zero Secrets: Building Secure Infrastructure for AI Agents
summary
Compares three concrete architectures — naïve host-privilege, MCP tool control, and sandbox + control plane — and argues infrastructure-level containment is required to limit agent blast radius. Demonstrates a local Docker Compose demo tying Anthropic completions, a GitHub MCP, MinIO (S3), and an OpenCode VS Code server into isolated containers that boot with a short-lived session token, sync artifacts to S3, and can be killed and resumed. Shows how scoped proxies and stateless sandboxes enable stateless disposable instances and central policy enforcement.
AI-written summary · about
chapters (7)
- 0:00AI agent security concerns
- 2:30AI agent misbehavior cases
- 8:06Defense layers for agents
- 16:54Secure agent design
- 21:05Live demo: secure architecture
- 28:55Key takeaways
- 31:50Q&A
related talks
39mTrust, but Sandbox: Securing AI Agents in Kubernetes
25mPlease Sandbox your AI Agents
11mBuilding Secure Backend Services with AI Agents
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
29mMCP Security: Keep Your AI Agents from Spilling the Tea
30mSandboxing Agentic AI With LSM-BPF
40mIn AI We Trust? Securing the Future, One Agent at a Time
44mMCP Security: Keep Your AI Agents from Spilling the Tea