Kubeshark patches two MCP server flaws that agents can be tricked into firing
Advisories published on 16 November 2026 describe an argument injection bug leading to Kubernetes cluster-admin takeover and an unvalidated file write, both fixed in v53.4.0.
By The Agentic Times ·

Kubeshark has patched two vulnerabilities in the Model Context Protocol (MCP) server built into its command line tool, according to two GitHub security advisories published on 16 November 2026. Both flaws affect versions v53.1.0 up to but not including v53.4.0, with v53.3.0 the last affected release. Both can be triggered by indirect prompt injection, where an AI agent ingests attacker-controlled content and replays it into a local tool call. The more serious of the two can end in full Kubernetes cluster-admin compromise. Both were reported through private coordinated disclosure and credited to a researcher using the handle novice-22 (Urim).
The first issue, tracked as CVE-2026-129236, is in the start_kubeshark MCP tool. According to the advisory, the tool built a kubeshark tap command by appending the caller-supplied pod_regex value as a bare positional argument, with no -- end-of-options separator. Because the underlying command parses flags using the cobra library, a pod_regex beginning with - or --set= was read as a flag rather than as a pod pattern. That allowed an attacker to inject Helm value overrides such as tap.docker.registry and tap.docker.tag, pointing the deployed worker, hub and front images at a registry they control. The worker runs as a privileged DaemonSet with host networking and the host filesystem mounted. In the reporter's proof of concept, the attacker image ran as root and read the host's cluster-admin kubeconfig files, the advisory said.
The tool is only registered when the operator starts the server with the non-default --allow-destructive flag. Kubeshark scored the issue 7.2 High, down from the reporter's 8.3, arguing the attack surface is local because kubeshark mcp is a stdio server with no network listener, and that the attacker borrows an existing high-privilege context rather than gaining one. The company did not contest the cluster-admin impact.
The second issue, CVE-2026-129235, is in the download_file tool, which is registered unconditionally and needs no special flag. The advisory says both the Hub API path and the local destination dest came straight from the tool arguments, with dest passed directly to file creation with no cleaning, no base-directory confinement and no rejection of absolute paths or ../ traversal. An attacker influencing those arguments could write fetched bytes to any path the user running the server can write, such as ~/.ssh/authorized_keys or ~/.kube/config, and could request arbitrary Hub endpoints. The reporter retrieved a captured-traffic PCAP file and recovered a planted bearer token from it. Kubeshark scored this 5.3 Medium, down from the reporter's 7.1, noting that reliable remote code execution was not demonstrated. It kept a high confidentiality rating because the Hub ships with authentication disabled by default on community installs, so captured traffic can be read end to end.
Both were fixed in v53.4.0 by the same commit. The pod_regex value is now appended last, after a literal -- separator, so it can never be parsed as a flag. For downloads, destinations are confined to a base directory, paths containing .. are rejected, and containment is re-checked after symlinks are resolved. Kubeshark's workarounds include avoiding --allow-destructive unless needed, running the server as an unprivileged user in an isolated working directory, restricting the kubeconfig available to that host, and not exposing the MCP server to untrusted agent input.
Sources
- github.comhttps://github.com/advisories/GHSA-8vf2-hxg5-crqh
- github.comhttps://github.com/advisories/GHSA-gj7r-p794-hhq5
