Summary

The OpenAPI adapter's spec-change poller (OpenApiSpecPoller) re-fetched the
configured spec url on a timer using a raw global fetch(), bypassing the SSRF
guard (safeFetch / assertUrlSafe) that OpenAPIToolGenerator.fromURL() applies
to the initial spec load. As a result, the pinning/DNS-resolution hardening delivered
via mcp-from-openapi >= 2.5.0 (advisory GHSA-65h7-9wrw-629c) protected the initial
load but not the recurring poll of the same URL. When polling is enabled against
an untrusted or attacker-influenceable spec URL, this is an unguarded SSRF vector.

Details

The initial spec load is guarded. OpenapiAdapter resolves a secure refResolution
policy and passes it to the guarded loader:

// libs/adapters/src/openapi/openapi.adapter.ts — initializeGenerator()
return await OpenAPIToolGenerator.fromURL(this.options.url, {
  // ...
  followRedirects: this.options.loadOptions?.followRedirects ?? false,
  refResolution, // secure default: external $refs off, internal targets blocked
});

But the poller — which re-fetches the same URL on every interval — did not:

// libs/adapters/src/openapi/openapi-spec-poller.ts — doFetch() (vulnerable, <= 1.5.5)
const controller = new AbortController();
const timeout = setTimeout(() => controller.abort(), this.fetchTimeoutMs);
try {
  const response = await fetch(this.url, {   // <-- raw global fetch, no SSRF guard
    headers,
    signal: controller.signal,
  });
  // ...hash the body, fire onChanged...
}

Because doFetch() never called safeFetch, none of the guard's protections applied
to the polled request:

• no allow-list / block-list enforcement (allowedHosts / blockedHosts);
• no internal/private/loopback/link-local/CGNAT/cloud-metadata IP blocking;
• no DNS resolution of the hostname (so a DNS name that resolves to an internal IP,
  e.g. http://127.0.0.1.nip.io/, was reached);
• no connection pinning to the validated IP (DNS-rebinding TOCTOU);
• no per-hop re-validation of HTTP redirects.

This is the identical threat model to fromURL() / external $ref resolution
(GHSA-65h7-9wrw-629c), applied to a request path that the fix for that advisory did
not cover.

Impact

A server that enables spec polling against an untrusted or attacker-influenceable
spec URL will, on every poll interval, issue a server-side GET to whatever host the
URL (or a DNS name it resolves to, or a redirect it returns) points at — including
internal-only addresses unreachable from the public internet. Consequences include:

• reading cloud-instance metadata endpoints (e.g. 169.254.169.254) — credential /
  token theft;
• probing and reaching internal services and private-range hosts (internal network
  scanning);
• DNS-rebinding to swap a public host for an internal one between validation and
  connection.

The poller issues GET requests only, so the primary impact is confidentiality
(reaching and reading internal endpoints); the fetched body is content-hashed to
detect change and the subsequent tool rebuild goes back through the guarded
fromURL() path.

Preconditions

Exploitation requires both:

1. polling.enabled: true on an OpenapiAdapter (polling is off by default and
   requires the URL-based url option, not an inline spec); and
2. the spec url is untrusted / attacker-influenceable (e.g. it is derived from user
   input, a tenant-supplied value, or otherwise not a fixed trusted constant), or an
   otherwise-trusted spec host is attacker-controlled or can redirect.

Servers that poll a fixed, trusted, first-party spec URL are not exposed in practice,
though they still benefit from the guard as defense-in-depth.

Proof of concept

import { OpenapiAdapter } from '@frontmcp/adapters';

// url is attacker-influenceable and points (directly, via DNS, or via redirect)
// at an internal target; polling re-fetches it every interval.
const adapter = OpenapiAdapter.init({
  name: 'evil',
  url: 'http://169.254.169.254/latest/meta-data/', // or http://127.0.0.1.nip.io/...
  polling: { enabled: true, intervalMs: 5000 },
});

await adapter.fetch();   // initial load IS guarded (blocked)
adapter.startPolling();  // <= 1.5.5: each poll issues an UNGUARDED GET to the internal target

On <= 1.5.5 the timed poll reaches the internal address. On the patched version the
poll fails closed (no request is made; the failure is logged) exactly as the initial
load does.

Patch

The fix routes the poller through the same SSRF guard as the initial load, with the
same policy, so both paths share one DNS resolution + connection pinning and cannot
diverge:

• OpenApiSpecPoller.doFetch() now calls safeFetch(this.url, { headers, timeoutMs, followRedirects, ssrf }) from mcp-from-openapi instead of the global fetch().
• OpenapiAdapter.startPolling() injects the adapter's resolved policy into the
  poller: ssrf: normalizeSsrfOptions(this.resolveRefResolution()) and
  followRedirects: loadOptions?.followRedirects ?? false — identical to what
  fromURL() receives.
• SpecPollerOptions gained optional ssrf / followRedirects; standalone use of
  OpenApiSpecPoller defaults to the secure policy (internal targets blocked,
  redirects not followed).

Files changed:

• libs/adapters/src/openapi/openapi-spec-poller.ts
• libs/adapters/src/openapi/openapi-spec-poller.types.ts
• libs/adapters/src/openapi/openapi.adapter.ts

Requires mcp-from-openapi >= 2.5.0 (already a dependency at 2.5.1), which exports
safeFetch / normalizeSsrfOptions and performs the resolved-IP validation and
connection pinning.

Remediation

Upgrade @frontmcp/adapters to 1.5.6 or later. No configuration change is required:
polling now inherits the same secure defaults as the initial spec load (external
targets blocked, redirects not followed). To poll a genuinely internal or localhost
spec server in a trusted environment, opt in explicitly with
loadOptions.refResolution.allowInternalIPs: true — the same knob that gates the
initial load.

Workarounds

For users who cannot upgrade immediately:

• disable polling (polling.enabled: false) on adapters whose spec url is not a
  fixed, trusted, first-party value; or
• only enable polling against spec URLs you fully control, served over HTTPS from a
  host that cannot be made to redirect to internal targets; and
• enforce network egress controls / an allow-list at the platform layer so the server
  cannot reach internal ranges or cloud-metadata endpoints.

References

• GHSA-8q49-2h5h-434x
• agentfront/frontmcp#510
• agentfront/frontmcp@077201e
• https://github.com/agentfront/frontmcp/releases/tag/v1.5.6