CVE-2026-46519

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.

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.

Summary

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.

Fixed in v3.6.0.

Impact

An attacker or misconfigured AI agent with network access to the MCP server's HTTP endpoint could invoke any Kubernetes tool regardless of the restriction mode configured by the operator -- including kubectl_delete, exec_in_pod, kubectl_generic, and node_management.

The project explicitly supports and documents multi-client HTTP deployment scenarios (Streamable HTTP and SSE transports, in-cluster deployments, Codex CLI and Gemini CLI integrations). In these deployments, operators relied on the tool restriction env vars to enforce least-privilege access across users or roles. The bypass invalidated that model entirely.

Severity scales with the Kubernetes service account's permissions. In environments where the MCP server runs with cluster-admin (common in dev/staging), this is equivalent to full cluster compromise for any client that can reach the endpoint.

The MCP_AUTH_TOKEN / X-MCP-AUTH mechanism controls who can reach the endpoint but provides no per-tool authorization. An authenticated client restricted to ALLOWED_TOOLS=kubectl_get could still invoke kubectl_delete after authentication.

Root Cause

In src/index.ts, the ListToolsRequestSchema handler applied the configured filtering logic before returning available tools. The CallToolRequestSchema handler dispatched directly by tool name with no equivalent check -- every tool was reachable unconditionally.

Proof of Concept

Tested across all three restriction modes against a live kind cluster. In each case, kubectl_delete was absent from tools/list but executed successfully via a direct tools/call request:

curl -s http://<HOST>:3003/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"kubectl_delete","arguments":{"resourceType":"pod","name":"test-pod","namespace":"default"}}}'

Result: {"result":{"content":[{"type":"text","text":"pod \"test-pod\" deleted\n"}]}}

Confirmed across ALLOW_ONLY_READONLY_TOOLS=true, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS=true, and ALLOWED_TOOLS=kubectl_get.

Remediation

The fix applies the same filtering logic from ListToolsRequestSchema at the start of the CallToolRequestSchema handler, returning an error for any tool call outside the active allowed set. Fixed in v3.6.0.

Credit

Discovered by Francisco Rosales of Manifold Security, coordinated by Ax Sharma, Head of Research at Manifold Security.

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.

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.

Summary

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.

Fixed in v3.6.0.

Impact

An attacker or misconfigured AI agent with network access to the MCP server's HTTP endpoint could invoke any Kubernetes tool regardless of the restriction mode configured by the operator -- including kubectl_delete, exec_in_pod, kubectl_generic, and node_management.

The project explicitly supports and documents multi-client HTTP deployment scenarios (Streamable HTTP and SSE transports, in-cluster deployments, Codex CLI and Gemini CLI integrations). In these deployments, operators relied on the tool restriction env vars to enforce least-privilege access across users or roles. The bypass invalidated that model entirely.

Severity scales with the Kubernetes service account's permissions. In environments where the MCP server runs with cluster-admin (common in dev/staging), this is equivalent to full cluster compromise for any client that can reach the endpoint.

The MCP_AUTH_TOKEN / X-MCP-AUTH mechanism controls who can reach the endpoint but provides no per-tool authorization. An authenticated client restricted to ALLOWED_TOOLS=kubectl_get could still invoke kubectl_delete after authentication.

Root Cause

In src/index.ts, the ListToolsRequestSchema handler applied the configured filtering logic before returning available tools. The CallToolRequestSchema handler dispatched directly by tool name with no equivalent check -- every tool was reachable unconditionally.

Proof of Concept

Tested across all three restriction modes against a live kind cluster. In each case, kubectl_delete was absent from tools/list but executed successfully via a direct tools/call request:

curl -s http://<HOST>:3003/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"kubectl_delete","arguments":{"resourceType":"pod","name":"test-pod","namespace":"default"}}}'

Result: {"result":{"content":[{"type":"text","text":"pod \"test-pod\" deleted\n"}]}}

Confirmed across ALLOW_ONLY_READONLY_TOOLS=true, ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS=true, and ALLOWED_TOOLS=kubectl_get.

Remediation

The fix applies the same filtering logic from ListToolsRequestSchema at the start of the CallToolRequestSchema handler, returning an error for any tool call outside the active allowed set. Fixed in v3.6.0.

Credit

Discovered by Francisco Rosales of Manifold Security, coordinated by Ax Sharma, Head of Research at Manifold Security.

These are all security issues fixed in the atril-1.28.4-1.1 package on the GA media of openSUSE Tumbleweed.

This update for atril fixes the following issues:

Changes in atril:

• Update to version 1.28.4 (bsc#1265880 CVE-2026-46519):

  • Build fixes
  • Fix tests imported from XReader
  • Fix tests with AT-SPI2 >= 2.53
  • Improve search system
  • pdf: Always use poppler_document_save to avoid data loss
  • Use properties for can-zoom-in and -out
  • libview: Allow zooming to the limits of the scale
  • shell: Fix Max zoom in UI

• Update to version 1.28.2:

  • epub: Disable thumbnailing sidebar
  • Fix .cbr mimetype
  • Wayland: stop segfaults on some systems
  • replace deprecated gtk_menu_tool_button_new_from_stock
  • libview/ev-document-model.c remove one more deprecation warning
  • replace ev_document_model_get_dual_page with

• Update to version 1.28.1:

  • Update CBR library to libarchive in README.md
  • ci: fix travis build failures caused by recent travis changes
  • Cleanup icons Makefile
  • icons: Include higher resolution icons
  • Updated translations.