Securing AI Agents in Your Cluster

This is for you if

• You have to approve or secure an AI agent that touches your cluster
• You built an agent and want to pass the security review
• You run Kubernetes and know RBAC, NetworkPolicy and kubectl
• You are interviewing for platform, SRE or AI-infrastructure roles

What's inside

• A toy Python agent on Ollama you attack and then harden, layer by layer
• The lethal trifecta and the OWASP LLM 2025 and Agentic 2026 threat models
• Direct and indirect prompt injection demos against your own lab agent
• A tool allowlist, a lockfile that diffs definitions, and per-tool credentials
• A least-privilege ServiceAccount proven with a kubectl auth can-i matrix
• Hardened pods enforced by a ValidatingAdmissionPolicy, plus gVisor
• Default-deny egress that blocks the cloud metadata endpoint
• Guardrails in code: argument validation, redaction, approval gates, budgets
• A tool-call audit trail correlated with the Kubernetes audit log
• A red-team harness that measures attack success rate, and a 60-second kill switch

After this guide you can

• Threat-model an agent with the trifecta and the OWASP agentic top ten
• Scope an agent's identity, tools and network so a fooled model does little
• Enforce isolation and egress with admission policy and NetworkPolicy
• Build guardrails, budgets and an audit trail that survive a compromise
• Measure your defenses with a red-team harness and drill incident containment
• Answer the agent-security questions now standard in platform interviews

Contents

1. The mental modelp. 4
2. Build a toy agent, then break itp. 12
3. Tools, MCP and the agent supply chainp. 21
4. Least privilege: giving the agent its own identityp. 28
5. Sandboxing: containing code and shell executionp. 33
6. Cutting the network leg: egress policyp. 38
7. The hardened agent: guardrails in codep. 42
8. Audit, tracing and detectionp. 47
9. Red-teaming: making the testing a routinep. 50
10. Incident response and governancep. 53
11. Troubleshooting playbookp. 57
12. Production checklist and design decisionsp. 59
13. Interview questionsp. 62
14. Cheat sheetp. 70
15. 30-day practice planp. 73