k8s-sec-agent

A Kubernetes security audit agent that runs as a declarative agent in kagent. It audits clusters against CIS security benchmarks using MCP tools, with built-in privacy-preserving pseudonymization so sensitive cluster data never reaches the LLM in plain text.

What it does

• Runs CIS Kubernetes Benchmark checks via kube-bench as in-cluster Jobs
• Scans all pods for security issues: privileged containers, running as root, missing resource limits, writable root filesystems
• Checks network policy coverage per namespace to find unprotected workloads
• Analyzes RBAC bindings for cluster-admin usage and overpermissive service accounts
• Executes arbitrary kubectl commands (with human-in-the-loop approval) for deep-dive investigations
• Produces a scored security posture report with prioritized findings and copy-pasteable remediation

How it works

The agent exposes five MCP tools through a FastMCP server. All tool outputs are pseudonymized before the LLM sees them — namespace names, pod names, IPs, service accounts, and secrets are replaced with scoped tokens (e.g., ns-a3f2-1, pod-a3f2-3). An OpenAI-compatible reverse proxy sits in front of the upstream LLM, rehydrating pseudonymized tokens in the response before it reaches the user. Each session gets its own sanitizer with a unique scope, ensuring multi-tenant isolation. Deployed via Helm chart into any kagent-enabled cluster.