ApostropheCMS: Host header controls server-side image fetch origin in @apostrophecms/ai-helper, enabling authenticated SSRF

Summary

The @apostrophecms/ai-helper package builds the URL for a server-side image fetch by resolving a storage-relative path against req.baseUrl. When the site does not set a global base URL (the default), req.baseUrl is derived from the incoming HTTP Host header. An authenticated editor can generate a base64 AI image (which is persisted without a stored url) and then send an unauthorized PATCH accept request with an attacker-chosen Host header. The server then issues an outbound GET to the attacker-selected origin, producing a Server-Side Request Forgery primitive that reaches hosts and ports the client cannot otherwise contact. The affected code lives in packages/ai-helper/modules/@apostrophecms/ai-helper-image/index.js.

Root Cause

The image accept route trusts req.baseUrl as a safe origin for constructing an internal fetch target, but req.baseUrl is attacker-influenced whenever the site has no configured base URL. The URL builder concatenates a default local storage path (which is relative, for example /uploads) and resolves it against req.baseUrl; because the left operand is relative, the resulting origin is taken entirely from req.baseUrl. Two protections that exist elsewhere are missing on this path: the generate route enforces an edit capability, but the accept route performs no authorization or ownership check, and no allowlist or validation constrains the resolved origin before the fetch.

packages/ai-helper/modules/@apostrophecms/ai-helper-image/index.js
  aiHelperImageUrl (line 91-92): image.url || new URL(self.uploadfs.getUrl() + `/ai-helper-images/${image._id}.png`, req.baseUrl)
  aiHelperFetchImage (line 64-65): fetch(self.aiHelperImageUrl(req, image))
  apiRoutes.patch['ai-helper/:_id'] (line 314-333): findOne({ _id }) only; no permission or ownership check
packages/apostrophe/modules/@apostrophecms/i18n/index.js
  setPrefixUrls (line 799-820): req.baseUrl falls back to `${req.protocol}://${host}`, host from the Host header when apos.baseUrl is unset

Impact

An attacker who holds a low-privileged editor account (the capability required to generate an AI image, which inserts a base64 record without a stored url and returns its identifier) can coerce the server into issuing outbound GET requests to an origin of their choosing by supplying a crafted Host header on the accept request. This yields SSRF: probing internal hosts and ports that are not reachable from the attacker's network position, and limited semi-blind exfiltration where an image-parseable response at the fixed path becomes a viewable attachment. The request path is fixed to /uploads/ai-helper-images/<id>.png, so only the origin (host, port, and http scheme) is attacker-controlled; arbitrary-path retrieval such as cloud metadata traversal is not reachable through this primitive. No integrity or availability primitive was demonstrated. Exploitation requires the default deployment where no global base URL is configured and local file storage is in use, together with a configured provider that returns base64 image data.

Proof of Concept

Preconditions:
  - ApostropheCMS site with @apostrophecms/ai-helper installed and a configured
    base64-returning image provider.
  - No global base URL configured (apos.baseUrl unset), default local uploadfs.
  - Attacker holds an account with edit capability on at least one content type.

1. Authenticate as the editor and generate a base64 AI image: POST /api/v1/@apostrophecms/image/ai-helper with body {"prompt":"x"}. The response contains the new record identifier <_id>. The persisted record has no url field.
2. Send the accept request with a malicious Host header:

PATCH /api/v1/@apostrophecms/image/ai-helper/<_id> HTTP/1.1
Host: 169.254.169.254
Content-Type: application/json

{"accepted": true}

3. The server derives req.baseUrl as http://169.254.169.254, builds the fetch target http://169.254.169.254/uploads/ai-helper-images/<_id>.png, and issues the outbound request.

Derivation confirmed against source logic:
  base64 record (no url) + Host: 169.254.169.254
    -> new URL("/uploads/ai-helper-images/<_id>.png", "http://169.254.169.254")
    -> server fetches: http://169.254.169.254/uploads/ai-helper-images/<_id>.png
  control (apos.baseUrl set to https://cms.example.com):
    -> fetch target uses https://cms.example.com (Host ignored)
  control (record stores its own url):
    -> aiHelperImageUrl short-circuits to the stored url (no SSRF)

Attack Chain

1. Exposure: The @apostrophecms/ai-helper-image module registers POST and PATCH routes under /api/v1/@apostrophecms/image/ai-helper, reachable by any authenticated session on a site that installed the package.
2. Control: The attacker controls the HTTP Host header on the PATCH request; the global locale middleware copies it into req.baseUrl because apos.baseUrl is unset (packages/apostrophe/modules/@apostrophecms/i18n/index.js:805,816-819).
3. Path: The attacker first generates a base64 image via POST, which inserts a record without a url (ai-helper-image/index.js:282-284); the returned identifier is then used in the PATCH accept request, which calls aiHelperFetchImage (:332).
4. Guard: The PATCH handler checks only req.body.accepted and record existence via findOne({ _id }) (:318,:322); it performs no capability check and no owner scoping, unlike the POST route which calls checkPermissions (:176). No validation constrains the resolved origin before the fetch.
5. Primitive: aiHelperImageUrl resolves the relative storage path against the Host-derived req.baseUrl and aiHelperFetchImage calls fetch on the result (:91-92,:64-65).
6. Result: The server performs an outbound GET to the attacker-chosen origin at a fixed path, granting internal host and port reachability and limited semi-blind response retrieval that the attacker's network position does not otherwise permit.

Bypass Evidence

The generate route's checkPermissions guard (ai-helper-image/index.js:176) is not present on the accept route, so no authorization protects the fetch. The following refutations were tested and defeated:

• Default local storage returns a relative URL: uploadfs.getUrl() yields /uploads because the local backend builds uploadsUrl = (apos.baseUrl || '') + prefix + '/uploads', so new URL(relative, req.baseUrl) takes its origin from req.baseUrl. Configuring a base URL or an absolute storage URL (S3/CDN) defeats the attack, which is why those are negative controls, not the default.
• The base64 record has no persisted url: insertOne(image) stores the record before image.url is assigned on the in-memory response object (:282-284 versus :294), so the PATCH handler re-reads a record with no url and the short-circuit in aiHelperImageUrl does not apply.
• Ownership and enumeration: GET and DELETE on these records are userId-scoped and cuid2 identifiers are unguessable, so the practical actor is the generating editor; this bounds the required privilege to a low-privileged authenticated user rather than an anonymous or cross-tenant attacker.
• Not a duplicate: the fix for the core pretty-URL SSRF advisory lives entirely in the separate @apostrophecms/file module and its stream proxy and never touches packages/ai-helper; the rich-text import SSRF advisory uses a different route with an explicit hostname allowlist. No advisory exists for @apostrophecms/ai-helper. This is a parallel unguarded sink, not an incomplete fix of either.

Affected Versions

• Ecosystem: npm
• Package: @apostrophecms/ai-helper
• Confirmed affected range: >= 1.0.0-beta.10, <= 1.0.0-beta.12
• Latest release checked: 1.0.0-beta.12 (npm registry, dist-tag latest)
• Fix status: not fixed

The vulnerable pattern was introduced in 1.0.0-beta.10: earlier releases through 1.0.0-beta.9 fetched the stored provider URL directly (fetch(url)) with no req.baseUrl resolution and no uploadfs-relative construction, so they are not affected by this pattern. Inspection of the published tarballs for 1.0.0-beta.10, 1.0.0-beta.11, and 1.0.0-beta.12 confirms the full chain (Host-derived base resolution, unguarded accept route, base64 insert without a stored url) is present and unchanged through the latest published release. This package publishes only on the prerelease (beta) channel, and the latest published version remains vulnerable.

Suggested Fix

Do not treat req.baseUrl as a trusted origin for a server-side fetch. Enforce the same authorization on the accept route that the generate route applies, and add an ownership or scope check so a record can only be accepted by a user permitted to act on it. Independently, constrain the fetch target: resolve the storage-relative path against a server-configured, trusted base (for example the configured storage base URL) rather than the request-derived req.baseUrl, and reject or allowlist any resulting origin before calling fetch. Ensure the remediation covers both the base64 (no stored url) representation and the stored-url representation, and any other caller that reaches aiHelperImageUrl or aiHelperFetchImage. As an interim mitigation without a code change, configure a global base URL so the resolved origin is fixed, and place egress controls on the server so outbound requests cannot reach sensitive internal hosts or metadata endpoints.

Reported by zx (GitHub: @manus-pi).