Cache-Control is one of the most frequently misunderstood HTTP headers, due to its overlapping and perplexingly-named directives. Confusion around it has led to numerous security incidents, and many configurations across the web contain unsafe or impossible combinations of directives. Furthermore, the interactions between various directives can have surprisingly different behavior depending on your browser.
The objective of this document is to provide a small set of recommendations for developers and system administrators that serve documents over HTTP to follow. Although these recommendations are not necessarily optimal in all cases, they are designed to minimize the risk of setting invalid or dangerous Cache-Control directives.
Safe for PII
Don't cache (default)
API calls, direct messages, pages with personal data, anything you're unsure about
When you're unsure, the above is the safest possible directive for Cache-Control. It instructs browsers, proxies, and other caching systems to not cache the contents of the request. Although it can have significant performance impacts if used on frequently-accessed public resources, it is a safe state that prevents the caching of any information.
It may seem that using no-store alone should stop all caching, but it only prevents the caching of data to permanent storage. Many browsers will still allow the caching of these resources to memory, even if it doesn't write them to disk. This can cause issues where shared systems may contain sensitive information, such as browsers maintaining cached documents for logged out users.
Although no-store may seem sufficient to instruct content delivery networks (CDNs) to not cache private data, many CDNs ignore these directives to varying degrees. Adding private in combination with the above directives is sufficient to disable caching both for CDNs and other middleboxes.
Static, versioned resources:max-age=n, immutable
The value to use for n depends upon the application, and is ideally set to a bit longer than the expected document lifetime. One year (31536000) is a reasonable value if you're unsure, but you might want to use as low as a week (604800) for resources that you want the browser to purge faster.
Infrequently changing public resources or low-risk authenticated resources:max-age=n
If you have public resources that are likely to change, simply set a max-age equal to a number (n) seconds that makes sense for your application. Simply using max-age will allow user agents to still use stale resources in some circumstances, such as when there is poor connectivity.
There is no need to add must-revalidate outside of the unlikely circumstance where the resource contains information that must be reloaded if the resource is stale.
For brevity, this only covers the most common directives used inside Cache-Control. If you are looking for additional information, the MDN article on Cache-Control is pretty exhaustive. Note that its recommendations differ from the recommendations in this document.
max-age=n (and s-maxage=n)
instructs the user agent to cache a resource for n seconds, after which time it is considered "stale"
s-maxage works the same as max-age, but only applies to intermediary systems such as CDNs
tells user agents and intermediates not to cache anything at all in permanent storage, but note that some browsers will continue to cache in memory
contrary to everything you would think, does not tell browsers not to cache, but instead forces them to check to see if the resource has been updated via ETag or Last-Modified
essentially the same as max-age=0, must-revalidate
forces a validation when cache is stale – this can mean that browsers will fail to use a cached resource if it is stale but the site is down
generally only useful for things like HTML with time-specific or transactional data inside
if max-age is set, must-revalidate doesn't do anything until it expires
indicates that the body response will never change
when combined with a large max-age, instructs the browser to not check to see if it's still valid, even when user purposefully chooses to refresh their browser
indicates that even normally non-cacheable responses (typically those requiring Authorization) can be cached on public systems, such as CDNs and proxies
recommended to not use unless you're certain, as it's probably better to waste bytes than to make the mistake of having a private document get cached on a CDN
indicates that caching can happen only in private browser (or client) caches, not on CDNs
note that this wording can be deceiving, as “private” documents are frequently cached on CDNs, with high-entropy URLs
documents behind authentication are an example of a good target for the private directive
instructs browsers to use cached resources which have been stale for less than n seconds, while also firing off an asynchronous request to refresh the cache so that the resource is fresh on next use
can provide a decent performance boost, as long as you're careful to avoid any issues where you require multiple resources to be fresh in a synchronized manner
browser support is still limited, so if you decrease max-age to compensate, note that it will affect browsers that don't yet support stale-while-revalidate
Common anti-patterns and pitfalls
Surveys of Cache-Control across the internet have identified numerous anti-patterns in broad usage. This list is not meant to be extensive, but simply to demonstrate how complex and sometimes misleading that the Cache-Control directive can be.
While there are times that max-age and must-revalidate are useful in combination, for the most part this is saying that you can cache a file but then must immediately distrust it afterwards even if the hosting server is down. Instead use max-age=604800, which says to cache it for a week while still allowing the use of a stale version if the resource is unavailable.
max-age=604800, must-revalidate, no-cache
no-cache tells user agents that they must check to see if a resource is unmodified via ETag and/or Last-Modified with each request, and so neither max-age=604800 nor must-revalidate do anything.
You still see these directives appearing in Cache-Control responses, as part of some long-treasured lore for controlling how Internet Explorer caches. But these directives have never worked, and you're wasting precious bytes1 by continuing to send them.
Expires: Fri, 09 April 2021 12:00:00 GMT
While the HTTP Expires header works the same way as max-age in theory, the complexity of its date format means that it is extremely easy to make a minor error that looks valid, but where browsers treat it as max-age=0. As a result, it should be avoided in preference of the far more simple max-age directive.
Not only is the behavior of Pragma: no-cache largely undefined, but HTTP/1.0 client compatibility hasn't been necessary for about 20 years.
fresh — a resource that was last validated less than max-age seconds ago
immutable — a resource that never changes, as opposed to mutable
stale — the opposite of fresh, a resource that was last validated more than max-age seconds ago
user agent — a user's browser, mobile client, etc.
validated — the user agent requested a resource from a server, and the server either provided an up-to-date resource or indicated that it hasn't changed from the last request
Since then, a number of additional assessments have done, including in October 2016, June 2017, and February 2018. All three surveys demonstrated clear and continual improvement in the state of web security. As a year has gone by since the last survey, it seemed like the perfect time to give the world wide web another assessment.
The overall growth in adoption continues to be encouraging, particularly the rise in the HTTPS and redirections to HTTPS. Overall, an additional 170,000 sites on the Alexa Top 1M now support HTTPS and about 190,000 of the top million websites have decided to do so automatically by redirecting to their HTTPS counterpart.
Subresource Integrity has also seen a sharp increase in uptake, as more and more libraries and content delivery networks work to make its usage a simple copy-and-paste operation. We've also see X-Content-Type-Options gain signicantly increased usage, particularly given that its usage enables cross-origin read blocking and helps protect against side-channel attacks like Meltdown and Spectre.
While the usage of Content Security Policy has continued to grow, it seems to be slowing down a bit. Tools like the Mozilla Laboratory make policy generation a lot easier, but it still remains extremely difficult to retrofit CSP to old and sprawling websites like so many of the top million.
Lastly, whether a result of policy changes in how the HTTP Strict Transport Security preload list is administered or some weird bug in my code, the percentage of the Alexa Top 1M contained in the preload list fell slightly. Oddly enough, of the 20,105 sites that set preload, only 5,540 of them are actually preloaded.
Mozilla Observatory Grading
Progress continues to be made amongst the Alexa Top 1M websites, but the vast majority still do not use Content Security Policy, Strict Transport Security, or Subresource Integrity. When properly used, these technologies can nearly eliminate huge classes of attacks against sites and their users, and so they are given significant weight in Observatory grading.
Here are the overall grades changes over the last year. Please keep in mind that what is being tested now isn't the same as what was being tested three years ago. An A+ in April 2016 was considerably easier to acquire than an A+ is now.
As a result, the bifurcation in scores likely indicates that more sites have decided to take web security seriously while others at the tail have fallen further into failure.
The Mozilla Observatory recently passed an important milestone of 10 million scans and has now helped over 175,000 websites improve their web security.
That's a big number, but I would love to see it continue to grow. So please share the Observatory so that the web can keep getting safer. Thanks so much!
Released in 1997, Magic: The Gathering - BattleMage was a real-time strategy game by Acclaim Entertainment. Published for the PlayStation and PC, its gameplay was bears little resemblence to the Magic we know today. Nevertheless, it is filled with an incredible amount of lore from early Magic history.
Due to its age and rarity – as well as the storyline's many branching paths – this lore was long-since considered lost to the Vorthos community.
Given Magic’s return to Dominaria, and BattleMage's significance in cards such as Time of Ice, I thought it best to crawl through BattleMage's code to extract the lore contained within.
I do hope you enjoy these texts, which are ordered as they appear in the game. Please contact me if you notice any mistakes in the editing. Thanks!
Since then, a number of additional assessments have done, including in October 2016 and June 2017. Both of those surveys demonstrated clear and continual improvement in the state of internet security. But now that tools like the Mozilla Observatory, securityheaders.io and Hardenize have become more commonplace, has the excitement for improvement been tempered?
Improvement across the web appears to be continuing at a steady rate. Although a 19% increase in the number of sites that support HTTPS might seem small, the absolute numbers are quite large — it represents over 83,000 websites, a slight slowdown from the previous survey's 119,000 jump, but still a great sign of progress in encrypting the web's long tail.
Not only that, but an additional 97,000 of the top websites have chosen to be HTTPS by default, with another 16,000 of them forbidding any HTTP access at all through the use of HTTP Strict Transport Security (HSTS). Also notable is the jump in websites that have chosen to opt into being preloaded in major web browsers, via a process known as HSTS preloading. Until browsers switch to HTTPS by default, HSTS preloading is the best method for solving the trust-on-first-use problem in HSTS.
Content Security Policy (CSP) — one of the most important recent advances due to its ability to prevent cross-site scripting (XSS) attacks — continues to see strong growth. Growth is faster in policies that ignore inline stylesheets (CSS), perhaps reflecting the difficulties that many sites have with separating their presentation from their content. Nevertheless, improvements brought about by specification additions such as 'strict-dynamic' and policy generators such as the Mozilla Laboratory continue to push forward CSP adoption.
Mozilla Observatory Grading
Despite this progress, the vast majority of top websites around the web continue not to use Content Security Policy, Strict Transport Security, or Subresource Integrity. As these technologies — when properly used — can nearly eliminate huge classes of attacks against sites and their users, they are given a significant amount of weight in Observatory scans.
As a result of their low usage rates amongst top websites, they typically receive failing grades from the Observatory. But despite new tests and harsher grading, I continue to see improvements across the board:
As 976,930 scans were successfully completed in the last survey, a decrease in failing grades by 2.9% implies that over 27,000 of the top sites in the world have improved from a failing grade in the last eight months alone. Note that the drop in A grades is due to a recent change where extra credit points can no longer be used to move up to an A grade.
Thus far, over 140,000 websites around the web have directly used the Mozilla Observatory to improve their grades, indicated by making an improvement to their website after an initial scan. Of these 140,000 websites, over 2,800 have improved all the way from a failing grade to an A or A+ grade.
When I first built the Observatory at Mozilla, I had never imagined that it would see such widespread use. 6.6M scans across 2.3M unique domains later, it seems to have made a significant difference across the internet. I couldn't have done it without the support of Mozilla and the security researchers who have helped to improve it.
Please share the Mozilla Observatory so that the web can continue to see improvements over the years to come!
I was recently writing some code for the Mozilla Observatory to store and interact with the HTTP status codes. As part of my code, I wanted to ensure that I would only store these status codes if they were an integer as per the HTTP/1.1 specification:
The status-code element is a three-digit integer code giving the
result of the attempt to understand and satisfy the request.
While it is easy to create test cases for conditions that don't satisfy this requirement, it is somewhat more difficult to determine how third-party libraries will handle HTTP requests that fall outside this constraint. I looked around the internet for websites to help me test weird status codes, but most of them only let me test with the known status codes. As such, I decided to add arbitrary HTTP status codes to my naughty httpbin fork, called misbehaving.site.
What I discovered is that the various browser manufacturers have wildly different behavior with how they handle unknown HTTP status codes. Here is what the HTTP specification says that browsers should do:
HTTP status codes are extensible. HTTP clients are not required to
understand the meaning of all registered status codes, though such
understanding is obviously desirable. However, a client MUST
understand the class of any status code, as indicated by the first
digit, and treat an unrecognized status code as being equivalent to
the x00 status code of that class, with the exception that a
recipient MUST NOT cache a response with an unrecognized status code.
…so what happens in reality?
Chrome's behavior is strange, but surprisingly not the strangest of the major browsers:
For negative status codes, Chrome always displays HTTP status code 200. For 0, it simply displays Finished instead of the actual status code. It otherwise simply reflects the status code, unless it exceeds 2147483647 (231-1), in which case it displays 2147483647.
Note that when exceeding 2147483647, it displays this error in the console, despite the page otherwise loading normally:
It actually took me quite a while to figure out Firefox's behavior. Let's take a look:
Status codes in Firefox are modulo 65536 (216), unless it works out to 0, in which it displays status code 200.
This works up to a certain point, when it starts to display different behavior:
Note how the status icon (blue dot, yellow triangle, etc.) is dependent on the first digit of the status code, once Firefox has finished interpreting it.
Safari only accepts status codes between 1 and 999. Should the status code fall outside that range, it reflects the entire HTTP request as plaintext, headers and all:
It also displays this error in the browser console. I'm not sure why, as the output is just JSON and there isn't any script on the page:
Note if you serve from localhost instead of a remote server, it displays a different error:
Not to be left behind, Edge also has some unusual HTTP status code handling:
For status code 0, it displays (Pending), although the page otherwise loads normally. For negative status codes, it displays them as the status code modulo 4294967296 (232). This is unless the status code is less than -4294967295, in which case it displays 1.
For positive status codes, it simply reflects them. This is until the status code reaches 4294967296 or higher, in which case it shows (Pending) and the browser displays this error:
Those who have been around in computing for a long time are likely familiar with Postel's Law:
Be liberal in what you accept, and conservative in what you send.
While it seems like the neighborly thing to do, it is the bane of those of us who enjoy consistent software behavior. If the specification had simply stated that status codes falling outside 100-599 should be treated as an unrecoverable error, then we wouldn't see the unusual behavior that we see today.
Luckily, while all of the browsers have their own idiosyracies, none of them are actually harmful in this case.
If you enjoyed this post and would like to test how browsers handle other quirky HTTP responses, please consider opening an issue or sending a pull request to the misbehaving.site github repository.