CVE-2026-46519: mcp-server-kubernetes Access Control Bypass – Patch to 3.6.0
mcp-server-kubernetes is a protocol server that helps manage Kubernetes clusters through a standardized interface. The software offers three environment variables designed to limit which cluster operations a user can perform—read-only mode, non-destructive operations only, or a custom whitelist. Before version 3.6.0, these restrictions only applied when listing available tools, but not when actually executing them. An authenticated user could bypass all restrictions by directly invoking a tool name they discovered, gaining full control over the cluster regardless of the configured safety mode. Version 3.6.0 fixes this enforcement gap.
Source data · NVD / CISA · public domain
- CVSS
- 3.1 · 8.8 HIGH · CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
- Weaknesses (CWE)
- CWE-863
- Affected products
- 0 configuration(s)
- Published / Modified
- 2026-06-11 / 2026-06-17
NVD description (verbatim)
mcp-server-kubernetes is a Model Context Protocol server for Kubernetes cluster management. Prior to version 3.6.0, mcp-server-kubernetes exposes three environment variables (ALLOW_ONLY_READONLY_TOOLS, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, ALLOWED_TOOLS) documented as access controls for restricting which Kubernetes operations are available. These controls are enforced at the tool discovery layer (tools/list) but not at the execution layer (tools/call). Any client that knows a tool name can invoke it directly regardless of the configured restriction mode. The access control was effectively cosmetic. This issue has been patched in version 3.6.0.
3 reference(s) · View on NVD →
SEC.co analysis · AI-assisted, reviewed against source
Technical summary
The vulnerability stems from a two-layer access control design where restrictions are validated inconsistently. The ALLOW_ONLY_READONLY_TOOLS, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, and ALLOWED_TOOLS environment variables are checked during tool discovery (the tools/list operation) but not during tool execution (the tools/call operation). This creates a privilege escalation path: an attacker with valid credentials can enumerate available tool names through discovery, then directly invoke restricted tools by name via the execution layer without passing through the access control checks. The flaw is classified as improper access control (CWE-863) and carries a CVSS 3.1 score of 8.8 (HIGH), reflecting the confidentiality, integrity, and availability impact possible through authenticated cluster manipulation.
Business impact
Organizations using mcp-server-kubernetes with role-based access controls face a critical trust boundary failure. If access restrictions are deployed to prevent junior developers, third-party integrations, or automated systems from performing destructive cluster operations, those restrictions provide no genuine protection. An attacker with any valid credential—including a low-privilege API token—can delete resources, modify configurations, disrupt services, exfiltrate data, or poison cluster state. This undermines compliance frameworks that rely on access controls and creates risk in multi-tenant or shared-cluster environments.
Affected systems
mcp-server-kubernetes versions prior to 3.6.0 are affected. All configurations using ALLOW_ONLY_READONLY_TOOLS, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, or ALLOWED_TOOLS environment variables are vulnerable, regardless of the specific restriction policy. Any deployment relying on these controls for security isolation is exposed.
Exploitability
Exploitation requires network access to the mcp-server-kubernetes endpoint and valid authentication credentials (PR:L in the CVSS vector). No complex attack chain is needed—a direct call to tools/call with a known tool name bypasses the restrictions entirely. The low complexity (AC:L) and lack of user interaction requirement (UI:N) mean exploitation is straightforward for any authenticated user aware of the tool interface. No special tools or knowledge of zero-day techniques are required.
Remediation
Upgrade mcp-server-kubernetes to version 3.6.0 or later, which enforces access controls consistently at both the discovery and execution layers. Review all cluster credentials and API tokens issued to users or systems relying on mcp-server-kubernetes access controls; assume those controls were ineffective before the upgrade and apply principle of least privilege during the remediation window. Verify that the patched version is in use across all deployments before releasing any sensitive cluster operations to untrusted users.
Patch guidance
Apply version 3.6.0 or any subsequent release. The patch closes the gap by validating the configured restriction mode before executing any tool, eliminating the bypass. Verify the patched version is running by checking the server version output and confirming the tools/call endpoint enforces the same restrictions as tools/list. Test in a staging cluster to ensure the restrictive behavior does not break legitimate automation before deploying to production.
Detection guidance
Monitor mcp-server-kubernetes audit logs and access patterns for calls to tools/call that would have been blocked by the restriction policy (e.g., destructive operations when ALLOW_ONLY_READONLY_TOOLS is set). Compare the tool names invoked via tools/call against the restriction rules to identify unauthorized operations. Ensure logging is enabled for all tool invocations. Look for authenticated sessions that use discovery (tools/list) followed by execution (tools/call) of tools that violate the stated policy, which may indicate awareness of the bypass.
Why prioritize this
This vulnerability rates HIGH priority due to the combination of high-impact blast radius (full cluster compromise possible), low exploitation friction (any authenticated user), and the prevalence of access controls in security-conscious deployments. Organizations using restrictive configurations are relying on controls that are completely ineffective, making immediate patching essential. The issue affects the integrity and availability of Kubernetes clusters, which are often critical infrastructure.
Risk score, explained
The CVSS 3.1 score of 8.8 reflects the HIGH severity: network-accessible service requiring only low-level authentication, low attack complexity, no user interaction, and unrestricted scope. The impact across confidentiality, integrity, and availability is complete (H:H/I:H/A:H) because an authenticated attacker gains full cluster control by executing any tool. This score is justified because the vulnerability nullifies the primary security mechanism organizations deploy on this server.
Frequently asked questions
Does this affect Kubernetes clusters themselves, or only the mcp-server-kubernetes tool?
The vulnerability is in mcp-server-kubernetes specifically. It does not indicate a flaw in Kubernetes itself. However, if mcp-server-kubernetes is used to manage your cluster and the access controls are bypassed, an attacker can cause damage directly to your Kubernetes cluster via that compromised interface.
What if we are not using the restriction environment variables at all?
If ALLOW_ONLY_READONLY_TOOLS, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, and ALLOWED_TOOLS are not configured, the server has no declared restrictions and the vulnerability does not apply. However, upgrading to 3.6.0 is still recommended for future consistency and to avoid accidental misconfiguration.
Can we mitigate this without upgrading, for example by network segmentation?
Network segmentation and firewall rules can reduce exposure, but they do not fix the underlying flaw. An attacker with valid credentials and network access will still bypass the restriction logic. Upgrade to 3.6.0 as soon as possible; do not rely solely on network isolation as a substitute for the patch.
What is the Model Context Protocol (MCP)?
MCP is an open standard that allows language models and AI systems to interact with external tools and data sources via a consistent interface. mcp-server-kubernetes is a specific MCP server that exposes Kubernetes cluster operations to AI-driven workflows or integrations. The restriction controls are intended to limit which operations such AI systems can perform.
This analysis is based on the official CVE record and vendor advisory. SEC.co does not independently verify all technical claims; users should consult the vendor's patch notes and security documentation before deploying. No exploit code or weaponized proofs-of-concept are provided. This information is current as of the publication date; refer to official sources for the latest patch status and compatibility details. Source: NVD (public-domain), retrieved 2026-07-20. Analysis generated by SEC.co (claude-haiku-4-5).
Weaknesses (CWE)
Related vulnerabilities
- CVE-2016-20075HIGHWordPress Ultimate Product Catalog 3.8.6 Arbitrary File Upload (CVSS 8.8)
- CVE-2025-14774HIGHABB T-MAC Plus Denial-of-Service Vulnerability (CVSS 7.4)
- CVE-2025-32348HIGHAndroid Local Privilege Escalation via Missing Permission Check
- CVE-2026-0272HIGHPalo Alto PAN-OS Privilege Escalation Vulnerability (PA-Series, VM-Series, Panorama)
- CVE-2026-21031HIGHAppBlock Authorization Flaw in Samsung Android—Risk & Patch Guidance
- CVE-2026-24724HIGHQNAP File Station 6 Authorization Bypass (CVSS 8.1)
- CVE-2026-3514HIGHPrefect 3.6.19 Authentication Bypass via Health Check Exemptions
- CVE-2026-35482HIGHalf.io Sandbox Escape Allows Admin Command Execution