While doing my research over the last months, I tried to come up with worst-case scenarios and payloads that could be employed against plain HTTP connections this way. Right from the start it had been evident that the Great Cannon could also be used to target individual users and organizations. The selectors for this targeting could be plain IP addresses or other user traits submitted via plain HTTP. HTTP injection is noisy however, especially since it has to be performed constantly against a user. For the Great Cannon this might be possible, but for an attacker trying to MiTM your HTTP connection, this is frequently not the case.
I realized that the scariest way to achieve this goal is by somehow persisting in the target browser after the injections have stopped. This way you could track users over extended periods of time, either as part of a botnet, to collect information, or to wait for the day that their browser or plugins are vulnerable to a remote-code-execution bug for a short period of time.
When thinking of persistence and HTTP, browser caching is the first thing that comes to mind. You could supply injected content along with custom caching headers which have a long cache duration. This will actually work reasonably well, but suffers from a couple of drawbacks:
The caching behaviour of different browsers is a complicated topic that I’m not gonna get into here. Suffice to say, major browsers nowadays offer easy ways of clearing the cache or hard-reloading a page.
My next idea was to go with the HTML5 Application Cache, aka the “offline website” functionality. You might have seen that in action when you used Google Docs and went offline: The web application is still available and fully usable.
The way that the AppCache works is by specifying a manifest file in the HTML header that tells the browser to download all of the files in the manifest and use them whenever the browser is determined to be offline. You can then make a website appear to be offline by DoSing either or cookie-bombing the user so the website appears to be offline. Both approaches make the actual website unreachable, which we want to avoid for obvious reasons.
Interestingly, the files in the AppCache are used even when the browser is online. In this case, the only thing the browser does is to check whether the manifest file changed. If it hasn’t it will happily keep using the offline files.
The obvious approach would be to inject a forged HTML, manifest and corresponding cache headers. Apparently, the cache headers are also applied to the manifest file. This is exactly what Egor Homakov did in his post on the Application Cache. But this approach suffers from the same drawbacks that pure cache-based attacks to: It can easily be circumvented by clearing the cache or hitting hard-refresh. For Safari, this will not work at all, since it will always check the manifest even if it had a long cache expiry time.
I started looking at the HTML5 spec and playing around with the Application cache. When you reload a page, the browser will try to get the manifest to see if it changed. If the manifest file is gone (HTTP 404 or 410), then the whole Application Cache will be cleared. This would defeat our one-time injection.
What caught my eye was this piece section:
Otherwise, if fetching the manifest fails in some other way (e.g. the server returns another 4xx or 5xx response or equivalent, or there is a DNS error, or the connection times out, or the user cancels the download, or the parser for manifests fails when checking the magic signature), or if the server returned a redirect, or if the resource is labeled with a MIME type other than text/cache-manifest, then run the cache failure steps.
W3C HTML5 spec
The cache failure steps finish with this step: “Abort the application cache download process.” To spell it out: The Application cache will stay intact if some part of updating it fails this way. The reasoning behind this is that the user might be behind some sort of captive portal (hence the 302), which effectively means “offline”. The same logic is true for return codes like 500 (“The server might be down”) or 200.
I only noticed the “Security Concerns” section in the draft spec after I had already come up with my attack. It addresses a very similar attack, but still does not quite apply to what we have in mind.
[…] an injection attack can be elevated into persistent site-wide page replacement. […] Targetted denial-of-service attacks or cookie bombing attacks can be used to ensure that the site appears offline. […] If a site has been attacked in this way, simply removing the offending manifest might eventually clear the problem, since the next time the manifest is updated, a 404 error will be seen, and the user agent will clear the cache. […] Unfortunately, if a cookie bombing attack has also been used, merely removing the manifest is insufficient; in addition, the server has to be configured to return a 404 or 410 response instead of the 413 “Request Entity Too Large” response.
W3C HTML5 spec
Now for the actual attack. Let’s take a step back and see what we’ve got:
In order to achieve our intended persistence, we need to inject the manifest on
a path that will generate a non-404/410 response from the legitimate
web-server! The path can be any sub-path, just make sure to deliver your fake
manifest file with MIME-type
The injected HTML at /:
The injected manifest file at
Picking a non-404/410 URL might sound complicated, but it’s actually fairly
simple: Pick any 302 or even 200 URL (as long as it does not reply with
This is where things get tricky. The question we want to answer is how persistent our injection really is. Remember that the AppCache functionality is supposed to work offline, so clearing the AppCache when the user hits “Reload” would defeat its purpose. Let’s have a look at each major browser:
Google Chrome will not clear the AppCache or reload the original HTML when
doing a refresh or hard-refresh (Ctrl-Shift-R). It will reload the original
file when opening the Inspector and selecting
[x] Disable Cache. Also, you
can examine and clear AppCaches in chrome via the internal URL
Firefox will not clear the AppCache when doing a refresh or hard-refresh. The only way to clear the AppCache is to go to Preferences -> Advanced -> Network and clear the AppCache manually.
Safari will not clear the AppCache on refresh or hard-refresh. On top of that, Safari will also not clear the AppCache even if you do Develop -> Empty Caches and Develop -> Disable Caches.
Safari does not expose the AppCaches in a very visible fashion. The only way I’ve found to clear an AppCache is by going to Preferences -> Privacy -> Remove All Website Data (or search for websites).
Opera will not clear the AppCache on refresh or hard-refresh (Shift+Refresh). Opera will clear the AppCache when hitting Clear Browsing Data.
Internet Explorer will not clear the AppCache on refresh or hard-refresh (Shift-Click). To clear it, you’ll have to go to Internet Options -> Settings -> Caches and databases.
Comparing the method to the Cache-Only attack of Homakov:
|Reload / Browse||Persist||Persist||Persist||Persist||Persist|
|Reload / Browse||Persist||Persist||Persist once||Persist||Persist|
As I said, browser caching is a complex topic. Suffice to say, employing only caching is not sufficient since the browser might clear the cache without a lot of user interaction. For example, Safari will clear the cache if you hit the regular “Reload” twice in a row.
There are a number of different scenarios where this attack can be used, ranging from stealthy to obvious.
The Great Cannon (or any large in-path system) could be leveraged to do these kind of injections. This would have to be very targeted, as any large-scale injection will immediately be noticed. Additionally, if someone already controls a system such as the Great Cannon, he would not really have to do these kind of injections to stay “persistent”.
Another scenario is much more dangerous: Local injections. If you’re going to a conference or browsing over a public hotel Wi-Fi, someone could MiTM you and thereby persist in your browser whenever you visit your favorite website. Personally, I browse a number of plain-only news websites which would be ripe targets as I go there daily and usually stay for a while to catch up on news.
An even more devious kind of injection would work against internal websites that the target might visit. This could be your average Intranet page which IT never protected with HTTPS since “it’s only reachable on the corporate network” anyway. Usually, the target would not visit these websites in a remote location, but he might have had an open tab or simply didn’t notice that the VPN connection had died again. In any case, you can now inject his internal website and grab whatever data you want the next time he goes online in his corporate network/VPN environment. Scary? It should be!
Both of these attacks could also be carried out without actually injecting the HTTP response. You could simply send a forged DNS reply to redirect the target to your AppCache-injection site once.
Lastly, as I mentioned, this kind of persistence would be ideal for actors without nation-state capabilities to pull of targeted exploitation. Without access to a treasure trove of 0days, you’d have to be lucky to catch your target in the narrow time-frame that he is exploitable. With a persistent injection, you could simply wait until there is yet another Flash/Java RCE 0day and exploit the user in this very instant.
Not really an injection, but this attack could just as well be mounted by an attacker who compromises the legitimate webserver and wants to persist even after his attack payload is cleaned from the server. In this case, the benign website could at least detect the injection and make sure it returns 404 for the fake manifest URL.
To prevent this attack from taking place in the first place, one should employ HTTPS along with HSTS whenever possible. HTTPS is not enough if users go to the plain HTTP URL: Deploy HSTS! There is a free CA launching this year, and the steps to actually get a proper HTTPS setup working have never been easier. If you’re in a corporate environment, you could set up your own CA and install its cert with your users. Make sure you understand the security-implications of this though! Custom CAs will not trigger key-pinning violations!
A mitigation if your website was attacked via an inject would be to change the requested paths for the faux manifest file to return 404 instead of some other return code.
Other things that could be done by Browser vendors would be to make the AppCache more visible to the user, via a dedicated icon in the address-bar that shows that the offline version of a page is in effect. Another feature or extension could be created which only allows AppCache to be set by HTTPS sites, which would get rid of the injection attack.
I had previously searched for existing documentation of this attack vector but could not find any mention of it. It was only after I had finished my research and this blog post that I came upon the blog-post by Attack&Defense Labs which describes the exact same attack :( It was then rehashed in this presentation by Krzysztof Kotowicz. I guess I only searched for “injection” rather than “poisoning”. Nevertheless, I still published this post as it should serve as a useful refresher and reminder and also examines different browsers and attack vectors.
As I’m not a cutting-edge kind of WebDev, I was not aware that AppCache will actually be replaced by ServiceWorkers. In the future, browser vendors might drop AppCache support, at which point it will be removed from the spec as well. Older browser will still support it though, and there is no timeline yet for when it will be removed.