@frontmcp/adapters
Affected
- introduced
0, fixed1.5.6(semver)
Fixed versions
1.5.6
Fix
- Fixed version
1.5.6npm install @frontmcp/[email protected]
Advisory
Also known as: None
FrontMCP: Server-Side Request Forgery (SSRF) in the OpenAPI adapter spec-change poller
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:
allowedHosts / blockedHosts);http://127.0.0.1.nip.io/, was reached);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:
169.254.169.254) — credential / token theft;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 guardedfromURL() path.
Preconditions
Exploitation requires both:
polling.enabled: true on an OpenapiAdapter (polling is off by default and requires the URL-based url option, not an inline spec); andurl 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.tslibs/adapters/src/openapi/openapi-spec-poller.types.tslibs/adapters/src/openapi/openapi.adapter.tsRequires mcp-from-openapi >= 2.5.0 (already a dependency at 2.5.1), which exportssafeFetch / 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 withloadOptions.refResolution.allowInternalIPs: true — the same knob that gates the
initial load.
Workarounds
For users who cannot upgrade immediately:
polling.enabled: false) on adapters whose spec url is not a fixed, trusted, first-party value; orMedium Basis: The source's own band: MODERATE
0, fixed 1.5.6 (semver)1.5.6
1.5.6npm install @frontmcp/[email protected]Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.
Check your own computer: guard packages check
Source: GitHub Advisory Database (reviewed advisories) · CC BY 4.0