/

CVE-2026-81091

Severity:

HIGH RISK

(8.6)

Server-Side Request Forgery (CWE-918) in mcp-use Proxy

CVE-2026-81091

Deferred

Vulnerability type

Category:

Server-Side Request Forgery

CWE:

CWE-918

Impact:

Unauthorized access to internal services and sensitive responses

Severity:

HIGH RISK

(8.6)

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 server-side request forgery vulnerability in mcp-use's Inspector proxy allows caller-controlled targets to determine where the server sends proxied requests. If an attacker can reach the proxy and provide a target URL, they may access loopback, link-local, private, or otherwise host-reachable services and read their responses, including through insufficiently validated redirects.

Affected context

The issue occurs in deployments using the mcp-use Inspector proxy when the destination is taken from the X-Target-URL header or the __mcp_target parameter. Risk increases when callers can reach the proxy endpoint and the server has network access to internal, loopback, link-local, or cloud-metadata addresses that are not available to the caller directly.

Exploit preconditions

  • Attack surface: The mcp-use Inspector proxy endpoint
  • Required access: Ability to send requests to the proxy and influence the target header or parameter
  • Trigger point: Proxy processing of a caller-supplied destination URL
  • Why exploitation is possible: Target hosts and resolved addresses were not adequately restricted before the request, and redirect destinations were not subjected to equivalent validation

Why this happens

The vulnerability results from treating a caller-controlled URL as an acceptable outbound destination without enforcing a network trust policy. The proxy accepted hosts that resolved to loopback, link-local, or private address ranges and did not reapply destination checks after redirects. This behavior maps to CWE-918: Server-Side Request Forgery, where an application makes network requests to attacker-influenced destinations from a trusted server environment.

Potential impact

  • Internal services may be accessed from the server's network position
  • Cloud metadata or other sensitive endpoints could be queried
  • Responses from otherwise inaccessible systems may be disclosed
  • Network segmentation controls may be bypassed
  • Redirect chains could extend access to additional protected destinations

General mitigation guidance

Mitigation generally focuses on enforcing an explicit outbound destination policy for proxy requests.

Common defensive patterns include:

  • Allowing only approved schemes, hosts, ports, and network destinations
  • Resolving hostnames and rejecting loopback, private, link-local, and other restricted address ranges
  • Rechecking the resolved destination immediately before connection establishment
  • Applying the same validation to every redirect, or disabling redirects unless explicitly required
  • Preferring an allowlist of known upstream services over caller-selected destinations
  • Restricting the proxy process's network permissions and limiting access to sensitive internal services
  • Avoiding disclosure of upstream responses or detailed connection errors when they are not required

How teams detect issues like this

Organizations can identify SSRF risks through:

  • Static analysis rules for HTTP clients that accept request URLs from headers, query parameters, or request bodies
  • Code review checks for unrestricted proxy targets, DNS-resolution gaps, and redirect handling
  • Tests covering loopback, private, link-local, reserved, and IPv6-mapped addresses
  • Network egress policies that restrict application access to internal and metadata services
  • Monitoring for unusual outbound requests, DNS lookups, redirect chains, and access to internal address ranges
  • Automated security review tools that flag user-controlled URLs passed to outbound request APIs

Sources

SourceLink
github.comgithub.com/mcp-use/mcp-use
github.comgithub.com/mcp-use/mcp-use/security/advisories/GHSA-f2jg-rm2x-hc5p
vulncheck.comvulncheck.com/advisories/mcp-use-inspector-proxy-server-side-request-forgery-via-caller-supplied-target-url

Frequently asked

questions

It may be remotely exploitable when the Inspector proxy is reachable by an attacker and accepts caller-controlled target values. The practical scope depends on authentication, exposure, and network placement.

The attacker generally needs permission to send requests to the proxy endpoint or access to another application feature that forwards input to it. They do not necessarily need direct network access to the internal destination.

Enforce a strict outbound destination policy using approved schemes, hosts, ports, and resolved addresses. Validate the destination before connection and repeat equivalent checks for every redirect.

Treat all proxy destinations as untrusted, use allowlists where feasible, apply network egress controls, test restricted address ranges, and monitor outbound traffic from proxy services.

Find out what's already
exploitable in your codebase.