- 1. GraphiQL introspection schema template injection attack: Advisory Statement
- 2. More Details on the Vulnerability
- 3. Compromised introspection Schema Example
This is a security advisory for an XSS vulnerability in graphiql
.
A similar vulnerability affects graphql-playground
, a fork of graphiql
.
There is a corresponding graphql-playground
advisory
and
Apollo Server advisory.
All versions of graphiql
older than
[email protected]
are
vulnerable to compromised HTTP schema introspection responses or schema
prop
values with malicious GraphQL type names, exposing a dynamic XSS attack surface
that can allow code injection on operation autocomplete.
In order for the attack to take place, the user must load a vulnerable schema in
graphiql
. There are a number of ways that can occur.
By default, the schema URL is not attacker-controllable in graphiql
or in
its suggested implementations or examples, leaving only very complex attack
vectors.
If a custom implementation of graphiql
's fetcher
allows the schema URL to be
set dynamically, such as a URL query parameter like ?endpoint=
in
graphql-playground
, or a database provided value, then this custom graphiql
implementation is vulnerable to phishing attacks, and thus much more readily
available, low or no privilege level xss attacks. The URLs could look like any
generic looking graphql schema URL.
Because this exposes an XSS attack surface, it would be possible for a threat actor to exfiltrate user credentials, data, etc. using arbitrary malicious scripts, without it being known to the user.
This advisory describes the impact on the graphiql
package. The vulnerability
also affects other projects forked from graphiql
such as
graphql-playground
and the graphql-playground
fork distributed by Apollo Server. The impact is
more severe in the graphql-playground
implementations; see the
graphql-playground
advisory
and
Apollo Server advisory
for details.
This vulnerability does not impact codemirror-graphql
, monaco-graphql
or
other dependents, as it exists in onHasCompletion.ts
in graphiql
. It does
impact all forks of graphiql
, and every released version of graphiql
.
It should be noted that desktop clients such as Altair, Insomnia, Postwoman, do not appear to be impacted by this.
[email protected]
addresses this issue via defense in depth.
-
HTML-escaping text that should be treated as text rather than HTML. In most of the app, this happens automatically because React escapes all interpolated text by default. However, one vulnerable component uses the unsafe
innerHTML
API and interpolated type names directly into HTML. We now properly escape that type name, which fixes the known vulnerability. -
Validates the schema upon receiving the introspection response or schema changes. Schemas with names that violate the GraphQL spec will no longer be loaded. (This includes preventing the Doc Explorer from loading.) This change is also sufficient to fix the known vulnerability. You can disable this validation by setting
dangerouslyAssumeSchemaIsValid={true}
, which means you are relying only on escaping values to protect you from this attack. -
Ensuring that user-generated HTML is safe. Schemas can contain Markdown in
description
anddeprecationReason
fields, and the web app renders them to HTML using themarkdown-it
library. As part of the development of[email protected]
, we verified that our use ofmarkdown-it
prevents the inclusion of arbitrary HTML. We usemarkdown-it
without settinghtml: true
, so we are comfortable relying onmarkdown-it
's HTML escaping here. We considered running a second level of sanitization over all rendered Markdown using a library such asdompurify
but believe that is unnecessary asmarkdown-it
's sanitization appears to be adequate.[email protected]
does update to the latest version ofmarkdown-it
(v12, from v10) so that any security fixes in v11 and v12 will take effect.
Note that if your implementation is depending on a CDN version of graphiql
,
and is pointed to the latest
tag (usually the default for most cdns if no
version is specified) then this issue is already mitigated, in case you were
vulnerable to it before.
If you cannot use [email protected]
or later
-
Always use a static URL to a trusted server that is serving a trusted GraphQL schema.
-
If you have a custom implementation that allows using user-provided schema URLs via a query parameter, database value, etc, you must either disable this customization, or only allow trusted URLs.
You can see an example on
codesandbox.
These are both fixed to the last graphiql
release 1.4.6
which is the last
vulnerable release; however it would work with any previous release of
graphiql
.
Both of these examples are meant to demonstrate the phishing attack surface, so
they are customized to accept a url
parameter. To demonstrate the phishing
attack, add ?url=https://graphql-xss-schema.netlify.app/graphql
to the
in-codesandbox browser.
Erase the contents of the given query and type {u
. You will see an alert
window open, showing that attacker-controlled code was executed.
Note that when React is in development mode, a validation exception is thrown
visibly; however that exception is usually buried in the browser console in a
production build of graphiql
. This validation exception comes from
getDiagnostics
, which invokes graphql
validate()
which in turn will
assertValidSchema()
, as apollo-server-core
does on executing each operation.
This validation does not prevent the exploit from being successful.
Note that something like the url
parameter is not required for the attack to
happen if graphiql
's fetcher
is configured in a different way to communicate
with a compromised GraphQL server.
This vulnerability was discovered by @Ry0taK, thank you! 🥇
Others who contributed:
- @imolorhe
- @glasser
- @divyenduz
- @dotansimha
- @acao
- @benjie and many others who provided morale support
The vulnerability has always been present
And later moved to onHasCompletion.js in 2016
(now .ts
after the typescript migration)
If you have any questions or comments about this advisory:
- Open an issue in graphiql repo
This section provides more details in addition to the advisory.
An installation of the GraphiQL web app is vulnerable if two conditions are met:
- The web app trusts information from its corresponding GraphQL server by interpolating information such as GraphQL type names directly into HTML instead of appropriately escaping or sanitizing the information.
- The victim can load the web app in a way where it speaks to a GraphQL server controlled by the attacker.
All versions of graphiql
prior to 1.4.7 inappropriately trust type names
provided by the GraphQL server. They additionally rely on XSS filtering in the
markdown-it
package to try to protect themselves from XSS attacks in GraphQL
descriptions and deprecation reasons.
By default, graphiql
does not allow the attacker to control what GraphQL
server it speaks to. Therefore, many installations of graphiql
are not
affected by this advisory. Installations are only affected if the fetcher
argument provided to GraphiQL allows arbitrary customization of the GraphQL
endpoint (eg, reading a GraphQL URL from an URL parameter), or if the attacker
has another way of affecting the introspection schema returned by the GraphQL
server. (Note that graphql-playground
, a project which started as a fork of
graphiql
, does this sort of URL parsing by default, so graphql-playground
installations are affected by a corresponding vulnerability in their default
configuration.)
One example of "another way of affecting the introspection schema" would be if
you served graphiql
as part of a PAAS platform that allows users to define
their own GraphQL schemas. In this case, even though the graphiql
installation
might be hard-wired to a single GraphQL endpoint, the attacker has control over
that GraphQL endpoint and could use it to inject scripts into graphiql
. In
this case, your graphiql
installation could be vulnerable if it responds to
introspection requests without first validating its schema. GraphQL servers can
prevent this by refusing to execute operations (including introspection
operations) on invalid schemas; any server built with graphql-js
properly
validates its schema prior to execution.
You can view the code for the exploited schema on codesandbox or in the repository
As you can see, the introspection schema must contain items with a compromised
name
value. this could be fields, input object names, enum names, variable
names, etc any graphql
NamedType
in the schema with it's name rendered in the autocomplete list.
{
"kind": "OBJECT",
"name": "<img src=x onerror=alert(document.domain)>",
"description": null,
"fields": [
{
"name": "name",
"description": null,
"args": [],
"type": {
"kind": "NON_NULL",
"name": null,
"ofType": {
"kind": "SCALAR",
"name": "String",
"ofType": null
}
},
"isDeprecated": false,
"deprecationReason": null
}
],
"inputFields": null,
"interfaces": [],
"enumValues": null,
"possibleTypes": null
}