What happened
Kubeshark's MCP tool start_kubeshark lacks an end-of-options separator when forwarding pod_regex, allowing Helm --set injection that redirects the container images kubeshark deploys. Because the worker DaemonSet runs privileged with host networking and host filesystem access, a malicious image can obtain node and cluster-admin access when the host can deploy it.
Why it matters
This is an MCP-tool argument-injection that turns a Kubernetes observability agent's MCP surface into a cluster-privilege-escalation path, illustrating the danger of passing untrusted agent-supplied arguments to privileged subprocesses.
Attack vector
The start_kubeshark MCP tool appends caller-controlled pod_regex as a bare positional argument to `kubeshark tap` without a `--` end-of-options separator (CWE-88), so a value beginning with --set is parsed as a Helm value override. This can redirect the privileged worker, Hub, and front images via tap.docker.registry/tag to an attacker-controlled registry; the privileged, host-networking worker DaemonSet can then yield node/cluster-admin access.
Affected systems
Kubeshark 53.1.0 through 53.4.0 (fixed in 53.4.0)
Mitigation
Upgrade Kubeshark to 53.4.0 (GHSA-8vf2-hxg5-crqh); avoid running kubeshark mcp with --allow-destructive unless necessary.