Kubeshark: MCP `start_kubeshark` argument injection via `pod_regex` → image redirect → cluster takeover
Ecosystem / package: Go — github.com/kubeshark/kubeshark (the kubeshark
CLI, kubeshark mcp subcommand)
Affected component: cmd/mcpRunner.go, MCP tool start_kubeshark
Affected versions: >= v53.1.0, < v53.4.0 (the flaw shipped with the MCP
command's first release; v53.3.0 was the last affected release)
Patched version: v53.4.0
Credit: novice-22 (Urim), via private coordinated disclosure.
Summary
The start_kubeshark MCP tool builds a kubeshark tap command by appending the
caller-supplied pod_regex as a bare positional argument, with no --
end-of-options separator. Because tap is a cobra command, a pod_regex that
begins with -/--set= is parsed as a flag, not as the pod regex. An
attacker who can influence the tool arguments (e.g. indirect prompt injection
against an agent driving kubeshark mcp --allow-destructive) can inject
arbitrary --set Helm value overrides — notably tap.docker.registry /
tap.docker.tag — redirecting the deployed worker/hub/front images to an
attacker-controlled registry. The worker DaemonSet runs privileged with
hostNetwork and the host filesystem mounted, so an attacker image yields node
and cluster-admin compromise.
Introduced in
Commit 0bbbb473 — "[1] Add MCP (Model Context Protocol) server command
(#1832)", 2026-02-06 — first released in v53.1.0 (v53.1.0~23). The
pod_regex argument has been appended as a bare positional since the tool's
introduction; there is no regression window, every MCP-capable release is
affected.
Details
In callStartKubeshark (cmd/mcpRunner.go), the command was built as
["tap", , "-n", …, "--set", …, "--set", "headless=true"] with
pod_regex inserted as a raw element. cobra parses flags positionally-agnostic
and there was no -- guard, so pod_regex="--set=tap.docker.registry=…,tap.docker.tag=…"
became a real --set. The tool is only registered when
kubeshark mcp --allow-destructive is set.
Proof of concept (reporter, abridged)
Driving kubeshark mcp --allow-destructive: start_kubeshark with
pod_regex="--set=tap.docker.registry=,tap.docker.tag=" caused
the CLI to construct kubeshark tap --set=… --set headless=true, deploying an
attacker image. The image ran as root inside the privileged worker container and
read the host's cluster-admin kubeconfigs
(/etc/kubernetes/super-admin.conf, admin.conf, pki/) — i.e. cluster-admin
takeover from a single pod_regex value.
Impact
Full Kubernetes cluster compromise (cluster-admin) from a single injected tool
argument, when the operator runs the MCP server with --allow-destructive.
Severity rationale
Kubeshark scores this 7.2 High, down from the reporter's 8.3, on two metrics. The
impact half of the vector (S:C/C:H/I:H/A:H) is unchanged — cluster-admin
takeover is real and Kubeshark does not contest it.
AV:N→AV:L.kubeshark mcpis a local stdio MCP server with no network listener. The attacker's payload reaches it as content an agent running on that host ingested and replayed into a local tool call. The poisoned content may have travelled over a network, but the exploited interface is local.PR:N→PR:H. Two administrative preconditions must already hold on the target host: the operator must have started the MCP server with the explicit, non-default--allow-destructiveflag (without itstart_kubesharkis not registered at all), and that host must already hold a kubeconfig with permission to deploy the privileged Kubeshark DaemonSet. The attacker is borrowing an existing high-privilege context, not gaining one.
Kubeshark kept AC:H (the injection must land in content the agent ingests, and
the attacker must stand up a malicious registry/image), UI:R (an operator
must be actively driving the agent), and S:C (the compromise crosses from
the CLI's authorization scope into the Kubernetes cluster — this is what keeps
the score in the High band).
Patches
Fixed in v53.4.0 by commit
a76a1920
(PR #1957).
In cmd/mcpRunner.go: pod_regex is now appended last, after a
literal -- end-of-options separator, so it is always parsed as the
[POD REGEX] positional and can never be interpreted as a flag. (namespaces /
release_namespace are passed through -n / -s value flags, which consume
their value token and are therefore not injectable.)
Workarounds
- Do not run
kubeshark mcpwith--allow-destructiveunless required. - Do not expose the MCP server to untrusted content / untrusted agent inputs.
- Restrict the kubeconfig/permissions available to the machine running the MCP server.