When you open a site, your browser receives two things. It receives a page, and it receives an envelope. The page is what you see on screen. The envelope is the HTTP headers, and it often contains information the site does not intentionally show, including the version number of the software running it. This number is not a detail reserved for IT specialists, it is what allows answering a very concrete question. Is your site exposed to a vulnerability that has just been published, or has it already fixed it. Without this number, the only honest answer is "we don't know", and an honest answer is better than a reassuring one. We apply this rule everywhere, and it explains why some sites in our estate carry the mention not assessable rather than a score. This guide shows the five places where a number can be read, with the exact commands, then what can be deduced and what must especially not be deduced. It requires no paid tool, no account, no installation, and it leaves no trace on the site you are looking at. A terminal is enough, on macOS, on Linux or on Windows, because the command used exists everywhere for over twenty years. Allow about ten minutes to go around your own site, and much less if you already have access to its administration.
A site is not one software, it is several
A website is a stack of several software, and each lives its own life. First there is the server that responds to your browser, and this server can be Apache, Nginx or IIS. Then there is the software that builds the pages, and this software can be WordPress, Joomla, Drupal or PrestaShop. There are the modules added on top, often about ten on a site that has been running for a few years, and that is where most of the actually exploited vulnerabilities hide. There is the theme that manages the appearance, and a theme abandoned for three years remains in place as long as no one replaces it. Finally there are the code libraries that run in your browser, such as jQuery or web fonts, and those have their own version numbers, visible to any visitor. Each of these pieces can have its own vulnerability, and that is what makes shortcuts dangerous. A site whose WordPress is perfectly up to date can be exposed by a module forgotten for three years, because no one thinks to update it, and the module remains active until it is uninstalled. The reverse is also true, a recent module placed on an old core protects from nothing. A site therefore does not have one version, it has one per software, and it is the oldest that decides the real risk. When we write that a site is exposed, we always speak of a specific software, with its version range, never of the site as a whole. The catalog that gathers these ranges can be consulted in the vulnerability records, and each record names the product concerned before giving the score and the share of our estate.
The five places where a number can be found
There is no single official source. A number can be read in five different places, and these places do not have the same reliability or the same availability depending on the site you are looking at. Some always respond, others are silent by choice, and a few lie through negligence when a partial update leaves an old file behind. The following order goes from the most direct to the most indirect, and it also corresponds to the order in which we ourselves read a site before writing its record. None of these readings modifies anything on the target site, they simply request a public page and look at the response, exactly as a visitor would. If one of them fails, it is not an error on your part, it is information about the site, and that information is as valuable as the others. A site that responds to three out of five sources is a talkative site, a site that gives none is a discreet site, and both can be perfectly up to date. What matters is not the number of sources that speak, it is what they say when they speak, and whether what they say stands on its own.
The Server header
curl -sI https://exemple.fr | head -3
HTTP/2 200
server: Apache/2.4.58 (Debian)
The first line of the envelope says which software is responding to your browser, and sometimes its exact number. Here the server announces 2.4.58, which is a direct reading, with nothing to deduce and nothing to assume. But many servers only respond Apache, without any number behind it, and then you have to understand why. This is neither a bug nor an oversight, it is a deliberate setting called ServerTokens Prod on Apache, recommended by all hardening guides. A server that does not give its number slightly complicates the task of an attacker looking for an easy target, and this discretion has a price, that of also preventing its owner from knowing where they stand. If your host has chosen to hide the number, the above command will tell you nothing more than the name of the software. There remains the possibility of asking for the number in writing, which is not the same as a reading, but which remains verifiable if the response is dated. We treat this response as a statement and not as a measurement, and we never record it in a record in place of a real reading, because a statement cannot be reproduced from one day to the next and can only be verified by hand.
The generator tag in the page
curl -s https://exemple.fr | grep -i generator
<meta name="generator" content="WordPress 7.1.4" />
Some software signs the pages it produces, by leaving a tag invisible on screen but present in the source code. It is found with a search in the raw text of the page, and this search does not require any particular tool. It is the software itself that declares its number, which makes it the most readable source that exists, and the easiest to interpret when starting out. The downside is that it is also the easiest to silence. A single line added to a configuration file is enough to make this tag disappear, and many sites remove it without knowing that they have just made their estate harder to audit. The tag sometimes also says something that is no longer true, when a site has migrated from WordPress to something else while keeping its old page templates, or when a theme module rewrites it. A generator tag is therefore not proof, it is a clue, and we treat it as such in our records. When it contradicts the server header, we keep both values instead of choosing one, because the contradiction itself is information about how the site is maintained.
The files of the technology itself
curl -s https://exemple.fr/wp-content/plugins/contact-form-7/readme.txt | head -5
The files of the technology itself carry the most stable trace of the entire inventory. Each WordPress module has a text file that begins with a Stable tag: line, each theme has a style file whose header contains a Version: line, and these files are public by construction. A very large number of sites leave them in place without a second thought, which makes them valuable. This source gives the module version even when the site has disabled all others, because the file is part of the module itself and is delivered with it. No one thinks to delete it during an update, so it remains in place for years. The file does not say whether the module is active, however, and that is an important limitation. A module that is present but disabled exposes nothing, so reading its number is not enough to conclude that a site is vulnerable, the module must also actually be loaded by a public page. We check this point before writing a record, because an alert sent for a disabled module is an alert that costs credibility for everything else. That is the difference between an inventory and a guess.
The ?ver= of core files
<link rel="stylesheet" href="/wp-includes/css/dashicons.min.css?ver=7.1.4" />
The ?ver= parameter does not always mean “site version”, and it is a trap we have fallen into. On a library address, the number is the version of that library, not that of the content management system. In a test conducted on thirty sites, a library number was read as a WordPress number, which would have produced a false comparison with the vulnerability catalog and an alert sent in error. Only a few core files actually carry the core number, and you have to know them one by one rather than accepting the first number that comes along. This lesson applies to all the readings that follow, a numerical value found in a page is a software version only if you know which software it refers to. The same caution applies to numbers that appear in source code comments, in dated file names, or in hand-written footers. An isolated number is not a measurement, it is a pebble, and it is the context that decides whether it belongs to the building, and that context requires two minutes of careful reading on the page itself, not on a screenshot.
The RSS feed and the OPML file
curl -s https://exemple.fr/feed/ | grep -o '<generator>[^<]*'
curl -s https://exemple.fr/wp-links-opml.php | grep -o 'wordpress.org/?v=[0-9.]*'
Two files that WordPress generates on the fly, on each request, contain the site version in plain text. The first is the article feed, the one feed readers use to follow a blog, and the second is a file inherited from an era when WordPress was used to maintain a list of links. Both are often forgotten in masking configurations, and that is precisely why they are useful. They still respond on a large portion of installed sites, they cost nothing to consult, and they leave no trace in the site logs. When the five sources contradict each other, which happens on poorly updated sites, we retain the oldest of the versions read, because a partial update almost always leaves one file behind. This is the prudent choice, and it sometimes designates a site as exposed when it no longer quite is. The opposite choice, retaining the most recent version, would produce the opposite error and would let truly vulnerable sites pass. Between two possible errors, we prefer the one that warns too much, as long as it is explained and dated. A site wrongly warned can respond, a site never warned will know nothing.
Why do so many sites say nothing ?
On a site served behind a content delivery network like Cloudflare, the real server never responds directly to your browser. It is Cloudflare that speaks, and Cloudflare gives no version number for the machines it protects. On a fully static site, there is no content management software, so no core number to read, and the absence of a number then means nothing worrying. On a site whose page is served from a cache, the envelope has sometimes been cleaned along the way, and the header you read is that of the cache, not that of the server. These three situations produce the same empty response although they have nothing in common from a risk perspective. To these must be added the most frequent case, that of a correctly configured site, where masking its number is a good practice recommended by hosting providers. A site that says nothing is therefore not a site in danger, it is a discreet site. The consequence for measurement is important, one can visit a site, observe that it is alive, and leave without any number. This is not the same as a site that has never been looked at, and it is not either a site known to be up to date. It is a site about which nothing can be said on this specific point, and this nuance changes everything when publishing a figure. Out of the 995,416 live sites in our registry, 1,518 are silent in this way, and the detail of this silence can be read in the state of the fleet as of November 13.
What we do with a number once read
A version alone says nothing at all. It only takes on meaning when compared to the bounds of a published vulnerability. A vulnerability is described by a range of affected versions, for example all versions prior to 2.4.59. If the site announces 2.4.58, it falls within the range and is exposed. If it announces 2.4.61, it is beyond the fix. The comparison is mechanical, reproducible by anyone, and verifiable line by line, which is exactly what is expected from a measurement. An approximate comparison, made on a library number or on a guessed version, produces an error that propagates to the person deciding whether or not to update. This is where everything is at stake, and this is what explains our insistence on reading. The difference between "this site may be vulnerable" and "this site is vulnerable" depends entirely on the quality of the version read. A recent example shows the mechanism in full, that of the Apache vulnerability of November 11, where 3,699 sites announced a version within the bounds and where 1,363 were still there two days later. The figure of November 11 was not false on November 11, it was simply dated, and re-reading it on the 13th gave a second figure that says something else. Publishing a count without its date amounts to presenting the first photograph as a permanent state, which is false even when the calculation is correct.
A version that cannot be read back is not a measurement. It is an impression.
When the site says nothing ?
When the five sources are silent, there remains a very strong temptation, that of completing. "It's WordPress, so it's probably a recent version." This sentence is not a measurement, it is an assumption, and it has two possible consequences depending on the direction in which it is wrong. If the assumption is too generous, a site that is truly vulnerable is cleared. No one will know, no one will fix it, and the site will remain exposed for months without its owner ever being warned. This scenario is the most frequent, and it is also the most costly in trust when it eventually becomes visible. If it is too severe, a site that has nothing is accused, which is defamation, and the site concerned is identifiable, so the harm is immediate. In both cases, the conclusion is not reproducible. Another day, another address, another cache, and the result changes without any software having moved an inch. A measurement that does not reproduce is not a measurement, it is an opinion, and an opinion cannot serve as a basis for a decision. This is why we write not assessable, and why we do not publish a vulnerability rate calculated on silent sites. A silent site is not an up-to-date site, it is out of reading range, and this sentence is enough to describe what we know.
What you can check yourself
Three commands are enough, and none of them modifies anything on the targeted site. You can run them from a terminal, on macOS, Linux or Windows, without installing anything other than the tools already present. If the three commands return nothing, your site is silent. That is not a problem in itself, and is even rather a good sign from a discretion standpoint. It is only a problem if you want to know, figures in hand, whether your site is up to date. In that case, the only way to know is to log in to the site administration and look at the updates page, which remains the only source that cannot be wrong about your own version. If your site is hosted by a provider, you sometimes do not even have access to that page. You then have to ask for the version in writing, with the date of the last security update, and keep the answer. A dated commitment is better than a reassuring version number, because it can be verified at the appointed time. It is the same logic that leads us to date every account published on this site, a number without a date does not say what is true, it only says what was true one day.
- Run
curl -sI https://votre-site.fr | head -5and look at the line starting withserver:. If a number appears after the software name, it is public. - Run
curl -s https://votre-site.fr | grep -i generator. If a version appears, your content software declares it itself. - If your site runs on WordPress, run
curl -s https://votre-site.fr/feed/ | head -20and look for thegeneratorline.
