/

CVE-2026-14537

Severity:

CRITICAL RISK

(9.8)

Incorrect Authorization (CWE-863) in MCP Toolbox API

CVE-2026-14537

Analyzed

Vulnerability type

Category:

Incorrect Authorization

CWE:

CWE-863

Impact:

Unauthorized Tool Invocation

Severity:

CRITICAL RISK

(9.8)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

A missing authorization check in the direct HTTP API tool invocation endpoint of Google MCP Toolbox may allow unauthenticated callers to invoke tools protected by the scopeRequired feature. The issue becomes risky when the legacy HTTP endpoints remain reachable while the --enable-api flag is active.

Affected context

The vulnerability occurs in Google MCP Toolbox deployments using the direct HTTP API invocation path. Exposure depends on the legacy HTTP endpoints being available, the API feature being enabled, and tools relying on scopeRequired authorization without an effective authorization check on that request path.

Exploit preconditions

  • Attack surface: Direct HTTP API tool invocation endpoints
  • Required access: No authentication may be required when the affected endpoint is reachable
  • Trigger point: A tool invocation request sent through a legacy HTTP endpoint while API access is enabled
  • Why exploitation is possible: The endpoint may fail to enforce authorization scopes before dispatching a protected tool

Why this happens

The underlying flaw is an authorization enforcement gap between the tool's declared protection requirements and the legacy HTTP request-handling path. A request can reach tool dispatch without an effective check that the caller possesses the required scope, which maps to CWE-863: Incorrect Authorization. The insecure pattern is trusting endpoint routing or feature configuration as a substitute for authorization at the protected operation.

Potential impact

  • Unauthenticated users may invoke tools intended for authorized callers
  • Sensitive data access or modification could occur through exposed tools
  • Administrative or operational actions may be performed without approval
  • Downstream systems could be affected depending on the permissions of the invoked tools

General mitigation guidance

Mitigation generally focuses on enforcing authorization consistently at every tool invocation boundary, including legacy and compatibility endpoints.

Common defensive patterns include:

  • Require authentication and validate scopes before dispatching any protected tool
  • Apply centralized, deny-by-default authorization middleware to all invocation routes
  • Remove or isolate legacy endpoints that are not required
  • Avoid treating feature flags or endpoint obscurity as access controls
  • Run tool services with least-privilege identities and narrowly scoped downstream permissions
  • Restrict network exposure and place administrative tool interfaces behind trusted access boundaries

How teams detect issues like this

Organizations can look for authorization gaps through:

  • Static analysis and code review of every route that dispatches tools
  • Tests verifying that unauthenticated and insufficiently scoped requests are denied consistently across current and legacy endpoints
  • Inventory and policy checks for exposed HTTP tool invocation interfaces
  • Logs showing unauthenticated requests, denied-scope failures, and successful tool executions
  • Automated API security testing that compares authorization behavior across equivalent routes

Sources

SourceLink
github.comgithub.com/googleapis/mcp-toolbox/pull/3435

Frequently asked

questions

It may be remotely exploitable when the affected HTTP endpoint is network-reachable, the API feature is enabled, and no additional access control blocks the request.

The described issue allows an unauthenticated attacker to invoke tools when the vulnerable legacy endpoint is exposed and fails to enforce the required scope.

Enforce authentication and scope authorization centrally before any tool dispatch, and apply the same checks to legacy, compatibility, and direct API routes.

Maintain deny-by-default authorization policies, test every invocation path for equivalent access control, minimize endpoint exposure, and apply least privilege to tool identities.

Find out what's already
exploitable in your codebase.