Cross-Origin Resource Sharing (CORS) is a mechanism which uses HTTP headers to tell a browser that a web application running at one origin has permission to access selected resources from a server at a different origin. This functionality exists for cases where an application developer would want to deliberately ignore a same origin policy (SOP) which mitigates many common attacks against browsers (notably cross-site scripting). This makes it a powerful tool for applications deployed on our increasingly distributed Internet but also a potential source of vulnerability.
Background: What is the Same-Origin Policy?
The Same-Origin Policy is a mitigating control which restricts how scripts or other resources from one origin interact with resources from a third party.
Some examples of cross-origin requests are:
- A different domain (example.com to test.com)
- A different subdomain (example.com to test.example.com)
- A different port (example.com to example.com:8080)
- A different protocol (https://example.com to http://example.com)
The Same-Origin policy mitigates some common web application attacks and is a critical tool for protecting users and applications on the modern web.
More details are available at the Mozilla Developer Network.
How Does CORS Work?
There are two headers which primarily govern CORS: Access-Control-Allow-Origin, and Access-Control-Allow-Credentials. When a script from foo.com wants to make a request to bar.com, the browser sends a pre-flight request to bar.com with foo.com in the Origin header. This is how the browser asks for permission for the resource to interact with the requesting site. The service on bar.com will then return an Access-Control-Allow-Origin response header if the domain matches an allowed origin. Only if the domain matches the one hosting the script does the browser send the HTTP request.
The Access Control Allow Origin header accepts various origins:
- Domains and subdomains (http://blah.example.com, https://example.com)
- Wildcards (*)
The case of the ‘Null’ origin is interesting and can result in some misconfigurations which we do not explore in this article. ‘Null’ is an origin assigned by default to many documents, sandboxed code, and redirects. While this is useful, it can potentially open up your application to attack. The Mozilla Developer Network and w3c specify that the null origin should not be used. More details about that vulnerability are available in Portswigger’s blog on the subject.
Access-Control-Allow-Credentials is a boolean – that is, it can be only True or False. If our application requires a user to be authenticated to use it, and that application makes CORS requests on behalf of that logged in user, we require Access-Control-Allow-Credentials to be true. If the resources we are accessing do not absolutely require that the user’s credentials be passed to the cross-domain resource, we should set this to False.
In some cases, you may want to bypass the same-origin policy to share content from a CDN, an S3 bucket, or an on-premises server.
- Loading a web font.
- Loading dynamic content from another webpage.
- Making an authenticated GET request to an on-prem server.
If you want to allow any domain to make a cross-origin request, you can certainly use the setting Access-Control-Allow-Origin: *
Unfortunately, per the CORS specification – if you have wildcards in your ACAO header, then Access-Control-Allow-Credentials cannot be true. While you might think that maintaining a list of allowed origins makes good sense, it turns out that CORS policy cannot accept a list either! So in order to allow credentials to be passed over a CORS request, you must process the origin and ensure that it’s valid. As we’ll explain, this is a recipe for disaster when developers are not careful.
CORS in the Cloud
In addition to the fact that any VM you deploy in your IaaS environment which supports HTTP can enable CORS, many individual services support CORS, including:
- Azure Functions
- Azure Logic Apps
- Azure Blob Storage
- Google Cloud Functions
- Google Cloud Storage
- Google Cloud Endpoints
- Amazon Lambda
- Amazon S3
- Amazon API Gateway
Clearly, this is a functionality that is widespread, and often used by web developers. As your infrastructure grows and matures, CORS is very likely to see use across your cloud infrastructure.
In the case of Lambda and API Gateway, Amazon provides the following guidance:
“To enable cross-origin resource sharing (CORS) for a proxy integration, you must add Access-Control-Allow-Origin: <domain_name> to the output headers. <domain_name> can be * for any domain name.”
As mentioned above, this works wonderfully, as long as you don’t need the Access-Controls-Allow-Credentials field to be true. If you need credentials, you’ll need to process the user-supplied origin. If you have never deployed CORS before and turn to the Internet for a template, you may end up using one of the many code examples which only reflect the origin and do not provide specific instructions on ensuring CORS is deployed securely.
Realistically, the most secure configuration is a hard-coded allow list that is maintained by the developer. However, as discussed below, even this does not stop CORS from opening up an exploitable hole.
One of the most common misconfigurations, as mentioned above, is reflecting the user-supplied origin header in the server response, effectively allowing requests from any origin. As an example, consider an AWS Lambda and API Gateway serverless architecture where the Lambda accesses data stored in a DynamoDB instance. If you want to include more than a single domain in the ACAO header (e.g., you have more than one subdomain that will need to access the API), then you’ll need to write a Lambda to handle the headers. Note that while we use AWS and Amazon services here as examples, these same sorts of vulnerabilities can also occur with the same architecture in both Azure and GCP. In the Azure case, this architecture would be comprised of Azure Application Gateway, Azure Functions, and CosmosDB. In GCP, the corresponding services would be Cloud Endpoints, Google Cloud Functions, and BigTable.
Figure 1: Simple CORS-enabled API Gateway Architecture
Now in this case, we can see that a request made to the API Gateway will trigger the lambda which generates the CORS headers. As we mentioned with this misconfiguration – the lambda is simply reflecting the origin provided by the user, and that request is then handed off to the second lambda to make database queries by way of the API gateway. This database can contain whatever sort of information you might access via API but don’t want disclosed to the world (e.g. customer records, API keys, etc.).
This means that all an attacker needs to do is get an unwitting but authorized user to make the request on their behalf from their domain. This can be attained by any number of methods: script injection, phishing, or any other way that a user might inadvertently navigate to an attacker-controlled resource. When the victim navigates to the attacker controlled resource, the sensitive request is made from the attacker’s origin with the authorized user’s credentials, and CORS allows this to occur, providing the sensitive data back to the attacker.
Impacts of CORS Misconfigurations
Examples of CORS misconfigurations being exploited:
- A US Department of Defense Website had an improper access control in CORS which allowed an attacker to steal user sessions.
- A bitcoin exchange had a vulnerability which could steal users’ private API key, allowing all of their BTC to be transferred to an arbitrary address.
These vulnerabilities and others like them underscore the need to verify that your CORS configuration is correct. This means ensuring that you are not simply reflecting the origin that you are provided by the browser, but maintaining an accurate, up-to-date allow list.
The Cream in Your XSS Coffee
(AUTHOR NOTE: It doesn’t matter if you personally like cream in your coffee. This is just a euphemism. I personally favor a venti Americano with a splash of cream and 3 splenda.)
Using CORS trust relationships, we can actually use even properly configured environments to make a cross-site scripting vulnerability on one site far more damaging. Given an XSS vulnerability on a page which is trusted via CORS, an attacker can retrieve sensitive information from the site which trusts the vulnerable page.
In this scenario, we’ll use the same API Gateway configuration as above, but assume that it actually does the CORS Origin checks correctly – that is, it looks for a valid subdomain and returns only if the user is authenticated and the request comes from a subdomain which is on an accurate whitelist.
Figure 2: Simple CORS-enabled S3 Front-End for API Gateway Architecture
These attacks and more have been detailed in a non-cloud context by the folks over at Portswigger, the makers of Burp Suite.
Even though these are hypothetical examples, knowing about CORS security is key because cross-domain requests are even more common in the cloud than they are in on-premises systems and so there are more opportunities for these sorts of vulnerabilities to exist and be exploited. So how do we deal with attacks that target our CORS-enabled cloud applications?
If you’re running cloud infrastructure, especially APIs in the cloud, you almost certainly have CORS deployed somewhere. Verifying that you do not have CORS misconfigurations in your cloud apps is a critical step in securing your cloud infrastructure. However, even if you have CORS properly configured, it can bite you if the apps which leverage CORS connections are not themselves secure since CORS can act as an amplifier for any latent vulnerability. By using the Mitre ATT&CK framework, we can create tabletop scenarios which allow us to profile and emulate attacks before they happen and give us a holistic view of our defense posture against certain types of exploits.