GHSA-65h7-9wrw-629c: SSRF
Summary
The published fix for GHSA-v6ph-xcq9-qxxj / CVE-2026-39885 added a direct hostname denylist for OpenAPI external $ref dereferencing, but the latest patched dependency mcp-from-openapi 2.3.0 still makes backend-origin requests to loopback when the target is reached through hostname resolution, redirects, or IPv4-mapped IPv6 syntax.
FrontMCP latest release v1.2.1 and current main still call OpenAPIToolGenerator.fromURL() and OpenAPIToolGenerator.fromJSON() from mcp-from-openapi 2.3.0 when loading OpenAPI adapters. An attacker who can cause a hosted or multi-user FrontMCP deployment to load an untrusted OpenAPI spec can trigger requests from the server to localhost or private services during tool generation.
This is a latest-version bypass of the previous fix. A direct http://127.0.0.1 $ref control is now denied and produces zero canary hits, while semantically equivalent loopback targets still reach the canary.
Latest versions checked
- frontmcp npm latest: 1.2.1 - @frontmcp/adapters npm latest: 1.2.1 - mcp-from-openapi npm latest: 2.3.0 - FrontMCP release tag: v1.2.1, commit db323976c66297d684a3e63bbfe1db6b310f2944 - FrontMCP current main checked: c15b79abe8c6a3cb71d4b7a3bafb8190730dc756
The release tag and current main both keep mcp-from-openapi 2.3.0 in package.json and libs/adapters/package.json, and both keep the OpenAPI adapter forwarding untrusted url, spec, and loadOptions.refResolution into OpenAPIToolGenerator.
Technical details
FrontMCP's OpenAPI adapter reaches the affected dependency paths:
- libs/adapters/src/openapi/openapi.adapter.ts imports OpenAPIToolGenerator from mcp-from-openapi. - loadOpenAPISpec() calls OpenAPIToolGenerator.fromURL(this.options.url, ...) and forwards loadOptions.refResolution. - The same method calls OpenAPIToolGenerator.fromJSON(this.options.spec, ...) and forwards loadOptions.refResolution.
In mcp-from-openapi 2.3.0, the patched guard is applied before the HTTP resolver fetches an external $ref. It checks the parsed URL hostname string against deny patterns for direct local and private addresses. The resolver does not resolve hostnames before allow or deny decisions, does not pin the validated IP to the fetch, and does not revalidate redirect targets before following them. It also misses IPv4-mapped IPv6 loopback forms.
As a result, these URLs are accepted by the guard but cause a loopback request from the backend:
- http://127.0.0.1.nip.io:<port>/schema.json, because the hostname string is not a direct IP even though it resolves to 127.0.0.1. - http://127.0.0.1.nip.io:<port>/redirect, because the first host passes and the actual request follows a redirect to http://127.0.0.1:<port>/schema.json. - http://[::ffff:127.0.0.1]:<port>/schema.json and http://[::ffff:7f00:1]:<port>/schema.json, because IPv4-mapped IPv6 loopback is not normalized and denied.
OpenAPIToolGenerator.fromURL() is also still unguarded for the initial OpenAPI spec URL. The PoC includes that as supporting evidence, but the primary report is the external $ref fix bypass.
Reproduction
The attached local PoC starts a loopback canary and loads generated OpenAPI specs using mcp-from-openapi 2.3.0. The request body schema contains a single external $ref for each test case. The canary records every backend-origin request.
Run:
bash cd /home/unkn0wn/securityaudit/frontmcp-ssrf-poc node repro-frontmcp-latest-ssrf-bypasses.mjs
Important output from a fresh run on 2026-05-25:
json {"name":"direct-127-denied-control","kind":"externalref","refUrl":"http://127.0.0.1:45117/schema.json","ok":false,"hitCount":0,"hits":[]} {"name":"dns-name-to-127-bypass","kind":"externalref","refUrl":"http://127.0.0.1.nip.io:45117/schema.json","ok":true,"hitCount":1,"hits":[{"url":"/schema.json","host":"127.0.0.1.nip.io:45117","authorization":null}]} {"name":"dns-name-to-127-bypass-with-allowedHosts","kind":"externalref","refUrl":"http://127.0.0.1.nip.io:45117/schema.json","ok":true,"hitCount":1,"hits":[{"url":"/schema.json","host":"127.0.0.1.nip.io:45117","authorization":null}]} {"name":"redirect-to-127-after-allowed-host","kind":"externalref","refUrl":"http://127.0.0.1.nip.io:45117/redirect","ok":true,"hitCount":2,"hits":[{"url":"/redirect","host":"127.0.0.1.nip.io:45117","authorization":null},{"url":"/schema.json","host":"127.0.0.1:45117","authorization":null}]} {"name":"ipv4-mapped-ipv6-dotted-bypass","kind":"externalref","refUrl":"http://[::ffff:127.0.0.1]:45117/schema.json","ok":true,"hitCount":1,"hits":[{"url":"/schema.json","host":"[::ffff:7f00:1]:45117","authorization":null}]} {"name":"ipv4-mapped-ipv6-hex-bypass","kind":"externalref","refUrl":"http://[::ffff:7f00:1]:45117/schema.json","ok":true,"hitCount":1,"hits":[{"url":"/schema.json","host":"[::ffff:7f00:1]:45117","authorization":null}]} {"name":"external-refs-disabled-control","kind":"externalref","refUrl":"http://127.0.0.1.nip.io:45117/schema.json","ok":true,"hitCount":0,"hits":[]}
Controls:
1. Direct loopback $ref is denied and the canary records zero requests. 2. Numeric IPv4 variants tested as parser controls were denied with zero requests. 3. Default file:// resolution was denied in this runtime. 4. Setting refResolution.allowedProtocols: [] prevents the external request, but this is not the default.
Impact
The previous advisory documented SSRF during untrusted OpenAPI $ref dereferencing. The latest patched version still lets an attacker trigger backend-origin requests to loopback or private network services through equivalent URL forms. In hosted or multi-user FrontMCP deployments where users can import or configure OpenAPI specs, this can expose internal admin APIs, metadata-like services, and other network endpoints that external users cannot reach directly.
The impact depends on whether a deployment treats OpenAPI adapter configuration as trusted administrator-only input. If untrusted authenticated users can import specs, this is a high-impact SSRF fix bypass. If only a local administrator can configure OpenAPI specs, the practical severity is lower.
Remediation
1. Do not rely on parsed hostname denylist checks for external $ref URLs. 2. Resolve hostnames before the request and reject loopback, private, link-local, multicast, unspecified, and metadata ranges. 3. Normalize IPv4-mapped IPv6 before range checks. 4. Revalidate every redirect target before following it, or disable redirects during external $ref dereferencing. 5. Pin the validated IP to the actual request with a custom dispatcher, lookup hook, or equivalent connect-time control. 6. Apply the same protected client to fromURL() initial spec loads. 7. Consider disabling external refs by default for untrusted OpenAPI specs and requiring explicit allowlists. 8. Add regression tests for direct loopback, DNS-to-loopback, redirect-to-loopback, IPv4-mapped IPv6, numeric IP forms, file refs, and disabled external refs.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
npm/frontmcpto a version that resolves this vulnerability.Fixed in 1.5.0 - Upgrade
Upgrade
npm/@frontmcp/adaptersto a version that resolves this vulnerability.Fixed in 1.5.0 - Upgrade
Upgrade
npm/mcp-from-openapito a version that resolves this vulnerability.Fixed in 2.5.0 - Upgrade
Upgrade
mcp-from-openapito a version that resolves this vulnerability.Fixed in 2.3.0 - Upgrade
Upgrade
frontmcpto a version that resolves this vulnerability.Fixed in 1.2.1 - Upgrade
Upgrade
@frontmcp/adaptersto a version that resolves this vulnerability.Fixed in 1.2.1 - Configuration
Set refResolution.allowedProtocols to [] for untrusted OpenAPI specs so external $ref dereferencing is prevented (disables external network requests during tool generation).
FrontMCP (OpenAPI adapter loadOptions.refResolution) refResolution.allowedProtocols = [] - Configuration
For untrusted OpenAPI specs, disable external refs by default and require an explicit allowlist (do not rely on parsed hostname denylist checks alone).
FrontMCP (OpenAPI adapter) loadOptions.refResolution = DISABLE_EXTERNAL_REFS_DEFAULT_FOR_UNTRUSTED - Configuration
Normalize IPv4-mapped IPv6 loopback representations before applying loopback/private/link-local range checks so forms like [::ffff:127.0.0.1] do not bypass SSRF protections.
mcp-from-openapi (external $ref dereferencing guard) ref resolution hostname/IP handling = NORMALIZE_IPV4_MAPPED_IPV6_BEFORE_RANGE_CHECKS - Configuration
Pin/validate the IP at guard time and ensure the HTTP client fetch uses the same validated address (e.g., via a custom dispatcher/lookup hook or equivalent connect-time control) rather than relying on hostname strings from the input URL.
mcp-from-openapi (external $ref dereferencing guard + fetcher) connect-time target pinning = PIN_VALIDATED_IP_TO_ACTUAL_REQUEST - Configuration
Resolve hostnames before allow/deny decisions for external $ref URLs, then reject loopback, private, link-local, multicast, unspecified, and metadata ranges based on resolved IPs.
mcp-from-openapi (external $ref dereferencing guard) DNS resolution before allow/deny decisions = RESOLVE_HOSTNAMES_BEFORE_REJECTING_LOOPBACK_PRIVATE_RANGES - Configuration
Revalidate every redirect target before following it during external $ref dereferencing; alternatively disable redirects during external $ref dereferencing so redirect-to-loopback cannot bypass the guard.
mcp-from-openapi (redirect handling for external $ref) redirect dereferencing behavior = REVALIDATE_EVERY_REDIRECT_TARGET_BEFORE_FOLLOWING_OR_DISABLE_REDIRECTS - Operational
Add regression tests covering direct loopback, DNS-to-loopback, redirect-to-loopback, IPv4-mapped IPv6 loopback forms, numeric IP forms, file:// refs, and disabled external refs to prevent recurrence of the SSRF bypasses.
Event History
Frequently Asked Questions
Which deployments are exposed to exploitation?
Hosted or multi-user FrontMCP deployments are exposed when an attacker can cause them to load an untrusted OpenAPI specification. The affected code path is used while loading OpenAPI adapters through OpenAPIToolGenerator.fromURL() or OpenAPIToolGenerator.fromJSON().
What does an attacker need to bypass the existing protection?
The attacker needs control over an OpenAPI specification that the server will load and can use external $ref targets that resolve to loopback or private services through hostname resolution, redirects, or IPv4-mapped IPv6 syntax. A direct http://127.0.0.1 $ref is denied, but semantically equivalent targets can still cause backend-origin requests.
Are the currently published package versions affected?
Yes. The advisory reports the bypass in FrontMCP 1.2.1, @frontmcp/adapters 1.2.1, and mcp-from-openapi 2.3.0, and states that FrontMCP current main was also checked.