TAU-HOME.COM
LOADING

Kubeshark MCP Server Security Vulnerabilities Disclosed: Cluster-Admin Takeover and File Overwrite Risks

Two security vulnerabilities in Kubeshark's MCP server—allowing cluster-admin privilege takeover (CVE-2026-129236) and arbitrary file overwriting (CVE-2026-1292

tau · November 21, 2026

#Kubeshark #MCP #AISecurity #Kubernetes #CVE

Kubeshark MCP Server Security Vulnerabilities Disclosed: Cluster-Admin Takeover and File Overwrite Risks

On November 17, 2026, two critical security vulnerabilities were disclosed in the Model Context Protocol (MCP) server of Kubeshark, a popular real-time network traffic analyzer and monitoring tool for Kubernetes clusters. These flaws highlight potential security risks when AI agents interface directly with privileged infrastructure tooling. The development team has responded with an immediate security patch in version 53.4.0.

Security advisory and code analysis interface regarding Kubeshark MCP server vulnerabilities

Image source: https://i10x.ai/news/cve-2026-129236-cve-2026-129235-kubeshark-mcp-server-flaws

The reported vulnerabilities comprise an argument injection flaw leading to cluster-admin privilege escalation, tracked as CVE-2026-129236 (CVSS score 7.2), and a path traversal and arbitrary file write issue, tracked as CVE-2026-129235 (CVSS score 5.3). Both vulnerabilities affected Kubeshark versions 53.1.0 through 53.3.0 and have been officially remediated in version 53.4.0.

Argument Injection and Cluster-Admin Takeover (CVE-2026-129236)

The first vulnerability, CVE-2026-129236, originated in how the start_kubeshark MCP tool handled command-line arguments within its codebase.

The tool appended a caller-supplied pod_regex parameter as a bare positional argument without a proper end-of-options separator. Because the underlying CLI parser evaluates flags positionally, an attacker capable of influencing tool arguments—such as through indirect prompt injection against an AI agent—could inject a value beginning with --set Helm override flags.

  • Attack Vector and Impact: By exploiting this behavior, an attacker could force Kubeshark to deploy worker and dashboard container images from an attacker-controlled registry rather than the legitimate one.
  • Preconditions: The attack required the operator to run the MCP server with the explicit --allow-destructive flag enabled and possess a kubeconfig holding sufficient deployment privileges. Security researchers demonstrated that meeting these conditions allowed attackers to read host cluster-admin configurations and achieve full cluster compromise.

Unvalidated File Paths and Traffic Data Exposure (CVE-2026-129235)

The second vulnerability, CVE-2026-129235, affected the download_file MCP tool due to insufficient path validation and default authentication configurations.

The tool accepted caller-controlled path and destination parameters, fetched requested assets from the Kubeshark Hub API, and wrote the response bytes directly to the local filesystem. However, the tool lacked proper filepath cleaning or base-directory containment checks, meaning absolute paths or directory traversal sequences (../) could write files to any location writable by the process.

  • Network Traffic Leakage: Furthermore, the Hub API answered without authentication by default in standard deployments, allowing captured traffic files (PCAPs) containing sensitive tokens, passwords, and credentials to be read and exfiltrated end-to-end.
  • Broader Attack Surface: Unlike the start_kubeshark tool, download_file was registered unconditionally without requiring the --allow-destructive flag, widening its exposure on default installations.

Security Recommendations and Upgrading

To mitigate these risks, the Kubeshark team has released version 53.4.0.

  • Fix for CVE-2026-129236: Appends a literal end-of-options separator (--) before the pod_regex argument, ensuring it is always parsed strictly as a positional parameter and never as an executable flag.
  • Fix for CVE-2026-129235: Introduces a new secureDownloadDest function that confines destination paths strictly to an allowable base directory, rejects traversal sequences, and adds proper credential forwarding and authentication checks.

Operators running versions 53.1.0 through 53.3.0 should upgrade to version 53.4.0 immediately. If an immediate upgrade is not feasible, administrators should avoid running the MCP server with --allow-destructive unless strictly necessary, execute it as an unprivileged user within an isolated directory, and ensure proper authentication is enabled on the Kubeshark Hub API.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285