We are considering to add the following feature to our web application (an online product database, if it matters):

  • Instead of uploading an image, the user can provide the (self-hosted) URL of an image. We store the URL instead of the image.

So far, so good. However, sometimes our web application will have to fetch the image from the (external, user-supplied) URL to do something with it (for example, to include the image in a PDF product data sheet).

This concerns me, because it means that our web server will send out HTTP requests to user-supplied URLs. I can immediately think of a lot of evil stuff that can be done with this (for example, by entering as the URL and trying out some common router web interface exploits). That seems similar to cross-site request forgery, only it's not the web server tricking the user into submitting a web request, it's the user tricking the web server.

Surely, I'm not the first one facing this issue. Hence, my questions:

  • Does this attack vector have a name? (So that I can do further research...)
  • Are there any other risks associated with fetching user-supplied URLs that I should be aware of?
  • Are there some well-established best-practice techniques to mitigate those risks?

3 Answers 3


This particular vulnerability indeed has a name. It is called Server-Side Request Forgery (SSRF). SSRF is when a user can make a server-side application retrieve resources that were unintended by the application developer, such as other webpages on an internal network, other services that are only available when accessed from loopback (other web services and APIs, sometimes database servers), and even files on the server (file:///etc/passwd). See the SSRF Bible and PayloadsAllTheThings for examples on how it can be abused. Since it's an image tag, most things probably won't be displayed, but it's still an issue to fix.

What to do about it? You can reference the OWASP SSRF Cheat Sheet. Your situation matches the second case, although you won't be able to perform all of the mitigations, like changing requests to POST or adding a unique token. The guidance otherwise boils down to:

  1. Whitelist allowed protocols: Allow HTTP and HTTPS, disallow everything else (e.g. a regex like ^https?://).
  2. Check that the provided hostname is public: Many languages come with an IP address library; check whether the target hostname resolves to a non-private and non-reserved IPv4 or IPv6 address*.
  3. My own addition, custom firewall rules: The system user that runs the web application could be bound to restrictive firewall rules that block all internal network requests and local services. This is possible on Linux using iptables/nftables. Or, containerize/separate this part of the application and lock it down.

Perhaps you could also validate the MIME type of the file at the time of retrieval to ensure it is an image. Also, do not accept redirects when fetching the image, or perform all the same validation on them if you do. A malicious webserver could just send a 3xx response that redirects you to an internal resource.

Additionally, you mentioned you are generating PDFs from user entered data. Your image URL aside, PDF generators have historically been a breeding ground for XXE (XML eXternal Entity injection) and SSRF vulnerabilities. So even if you fix the custom URL, make sure your PDF generation library avoids these issues, or perform the validation yourself. A DEFCON talk outlines the issue (PDF download).

* As mentioned in comments, DNS responses can contain multiple results, and responses could change between requests, causing a time-of-check time-of-use (TOCTOU) problem. To mitigate, resolve and validate once, and use that originally validated IP address to make the request, attaching the host header to allow the correct virtual host to be reached.

  • Also don't allow the user to choose the filename. Verify the mime type. Scan it for viruses (for several months after uploading).
    – symcbean
    Commented Apr 25, 2020 at 22:47
  • 26
    So you link to a PDF that talks about evilized PDFs? Hm ... Commented Apr 26, 2020 at 17:17
  • 1
    "Many languages come with an IP address library; check whether the target hostname resolves to a non-private and non-reserved IPv4 or IPv6 address." ideally you should resolve the hostname to an IP, validate the IP and then use the validated IP to make the request. Remember than DNS resoloution of a hostname may produce multiple results and may produce different results on different queries. Commented Apr 26, 2020 at 20:22
  • 2
    At this point, I wouldn't even allow HTTP external links. 90% of internet traffic these days goes over encrypted connections. There shouldn't be any reason to allow these images to be loaded over insecure protocols, especially with how cheap HTTPS integration is these days.
    – Nzall
    Commented Apr 27, 2020 at 13:15
  • 1
    @Tyzoid TLS overhead is minimal, especially if using HTTPS allows the use of HTTP/2.0. I'm not sure how many people are going to view images using IoT devices. There are definitely ways to cache items using HTTPS connections. And supporting old devices is a business decision, not a security decision. From a security perspective supporting old devices is a danger.
    – Nzall
    Commented Apr 28, 2020 at 10:00

We store the URL instead of the image.

In addition, this will add information and privacy risks. Let me show with a visual demo.

If you try to upload any image to StackExchange, you will notice that the image gets hosted by imgur.com. The SE server fetches the images and uploads a copy of it to its private server.

I will use a popular and innocent meme for the experiment. Let's start with the following URL for our show: https://i.imgflip.com/2fv13j.jpg. Plase note that I have to use a deep link for this demo to work.

I want to attach it to this post using StackExchange upload tool. Exactly like the scenario in the question.

A random image from the internet

🔝 Here is our newly upload image!

Let's go deeper and investigate futher. Notice that the image is now sourced from imgur.com rather than from imgflip.com. Please be patient with me if the two URLs have similar names. By opening Developer tools, you can see where the image is pointed

A screenshot from Developer tools

Privacy concerns

When you link just any http(s):// online resource, your browser will initiate a connection to that server, sending a lot of information. On a high-traffic website, the owner of the website gets a lot of information about IP addresses of people who visit this Security SE page, along with (if enabled) referral links and third party cookies. Considering that 3rd party cookies are enabled by default, this may leak the user identity if abused the right way.

By owning the image I want to upload to a post, StackExchange prevents imgflip.com from knowing who is displaying their picture.

And, as we are going to see in the second part, to change it in the future.

Risk of deception

Consider that no matter your effort to deploy a static "Front-page-ish" simple website, any URL to a remote resource is always interpreted by the server, on every request. While it may end with .jpg the server may likely be using a scripting language to interpret the request and, in the end, choose what content to serve.

Now that you have profiled your visitors, you have the power to choose what content to display for them, live. Consider the Uber-Greyball case as an example of live deception. Popular dating app Tinder uses a similar soft-ban or greylisting technology

Unknown to [...] authorities, some of the digital cars they saw in the app did not represent actual vehicles. And the Uber drivers they were able to hail also quickly canceled. That was because Uber had tagged [... Mr Police officer ...] — essentially Greyballing them as city officials — based on data collected from the app and in other ways.

As an example, the server can implement such a logic: decide where to serve an innocuous meme or an undesirable content, e.g. political advertising, based on the user requesting (reminds some Cambridge Analytic-thing?). Neverthles, the URL never changes.

CA's professed advantage is having enough data points on every American to build extensive personality profiles, which its clients can leverage for "psychographic targeting" of ads

request for https://host.com/images/img1.png
if (request comes from any of
      StackExchange moderator or
      Automated content filter or
      Government enforcer or
      Anything you can imagine)
    decide to serve innocuous content
else if (request comes from a user you want to decept)
    decide to serve a targeted advertising or deceptive content somehow

Look at this picture to see what may happen with real time filtering. A the same URL, different users see different content. Thanks to Lord Buckethead for keeping myself politically neutral.

Different content served from the same URL

At the same URL, we are now able to serve content that is different with regards to who is requesting it.

For these reasons, you have to consider fetching the remote resource in order to take a permanent snapshot of it, with regards to bandwidth and disk space constraints.

I won't discuss here about 1) retaining EXIF tags and 2) re-encoding the image with your own codec to prevent further payload attacks


Instead of uploading an image, the user can provide the (self-hosted) URL of an image. We store the URL instead of the image.

You mean these kind of JPEGs?

It is a bad idea. First of all, you will have to check the validity of the image every time you use it. That takes time. I assume that the database is used by other users, and you will have no control whatsoever what kind of malicious JPEGs the user of the database gets served. You are concerned that you get a malicious image, but you are willing to let others that use your database just get such a malicious image.

So, not so far so good.

For yourself, treat the image as you would treat any input from an untrusted source. That means: check that the image is correct. You might want to convert it to some standard format; you might want to re-encode to be sure.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .