Executive brief
mcp-server-kubernetes is a tool that allows AI agents to manage Kubernetes clusters. A security flaw in how it handles commands allows an attacker to trick the system into sending sensitive administrative login tokens to a server controlled by the attacker. If an operator uses an AI assistant to view logs or manage the cluster, the attacker can steal their credentials and gain full administrative control over the Kubernetes environment.
Technical details
The kubectl_generic tool in mcp-server-kubernetes fails to validate or allowlist user-supplied flags before passing them to the kubectl binary. An attacker can use indirect prompt injection (e.g., by planting malicious instructions in application logs) to influence an AI agent's call to this tool. By injecting the --server and --insecure-skip-tls-verify flags, the attacker can redirect kubectl API requests to a malicious HTTPS endpoint. Because kubectl sends the Authorization: Bearer token over HTTPS even with self-signed certificates, the attacker can capture and replay the operator's high-privilege credentials against the legitimate Kubernetes API. This issue is fixed in version 3.7.0.
Affected products
- Flux159 mcp-server-kubernetes <= 3.6.2
Timeline
- 2026-05-20: patched: Version 3.7.0 released
- 2026-05-22: advisory: GitHub Security Advisory published
- 2026-06-11: disclosed: CVE published to NVD