Secure websites shun HTTP Public Key Pinning

The HTTP Public Key Pinning header, or HPKP, can prevent fraudsters using mis-issued TLS certificates. While it offers a robust defence against website impersonation, hardly any HTTPS websites are actually making use of this powerful security feature, even though it has been supported by some browsers for more than a year.

Less than 0.1% of certificates found in Netcraft's March 2016 SSL Survey were served with the HPKP header. Where it has been deployed, a third of webmasters have mistakenly set a broken HPKP policy. With so many mistakes being made, the barrier to entry is evidently high.

Even for those webmasters who have set a valid policy, a lot of ongoing care and attention is required: both routine and emergency maintenance poses a significant risk of blocking legitimate visitors, potentially for long periods of time. However, when correctly deployed and carefully maintained, HPKP is a powerful security feature.

What does HPKP defend against?

A website can defend against most man-in-the-middle attacks by deploying HTTPS, HSTS and HSTS preloading. Together, these ensure all communication to and from the website is authenticated and encrypted.

While these provide a fairly robust defence against attacks like pharming and sslstrip, there is still a line of attack open. A knowledgeable and dedicated attacker can still attack an otherwise well-defended HTTPS website if he can convince a certificate authority to fraudulently issue him a certificate for it.

HPKP can prevent a customer from accessing a spoof website even if it uses a fraudulently-issued (but otherwise valid) certificate. Both sites use certificates issued by trusted CAs, but only the legitimate site's certificate chain contains a certificate that is expected by the client browser.

HPKP can prevent a customer from accessing a spoof website when it uses a fraudulently-issued (but otherwise valid) certificate. Both sites use certificates issued by trusted CAs, but only the legitimate site's certificate chain contains a certificate that is expected by the client browser. Hashes that correspond to each pinned certificate will have been noted by the browser during a previous visit to the legitimate site.

Although it is extremely difficult for a fraudster to obtain a certificate for a domain he does not control, it is not impossible. In fact, there is ample precedent. Several certificate authorities have been breached, lax issuance policies have been discovered, and technical flaws have been exploited by attackers.

The HPKP header is motivated by the history of mis-issuance within this ecosystem. To use HPKP, website owners must select a set of public keys that must be used in future connections. After visiting the site, its HPKP policy is then stored by the client to reject future connections to servers that use different, non-whitelisted keys.

However, creating an HPKP policy is not entirely sufficient to defend against impersonation attacks. In particular, HPKP cannot defend against rogue root certificates installed locally on users' computers.

Both Dell and Lenovo have recently been caught deploying local root certificates to their customers, along with accompanying private keys. With this knowledge, an attacker can generate a certificate for any website and use it to impersonate that site. The victim's browser will regard the certificate as valid, regardless of the genuine site's HPKP policy.

How is HPKP used?

There are three types of key that can be pinned using HPKP:

  • The current public key of the certificate issued to a site.
  • Public keys corresponding to certificate authorities and their intermediate certificates.
  • Backup keys.

In order for browsers to accept and store a website's HPKP policy, there must be at least two pins specified. At least one pin must be in the chain of trust formed by the browser when verifying the site's certificate, and there must be at least one pin that is not in the chain (a backup pin).

Here is an example of a valid HPKP header, which sets pins for three distinct public keys (marked in bold). This policy is valid for one year over all subdomains of the current origin:

  max-age=31536000; includeSubDomains

Webmasters must be cautious when pinning certificate authority keys. CAs may change their issuance practices without notice, and new certificates may not use the same chain of trust as the old ones. If the new certificate chain no longer includes the pinned keys, the website will not be accessible until the HPKP policy expires.

To avoid the problems posed by using certificate authority keys, webmasters can elect to pin their own keys. This is also a risky practice if the backup key cannot be used: it may have been lost, or may no longer qualify for inclusion in certificates (for example, if a backup key is known to be a Debian weak key, CAs will not accept it for use in new certificates).

"Who dares pins"

HPKP is perfectly safe to implement when pins and certificates are well-managed, but it can also be considered rather risky when you think about what could go wrong: A small mistake could effectively wipe out an online business by preventing its own customers from accessing its website for months. Here are some of the most popular sites that are brave enough to be using HPKP today:


GitHub is the busiest site to have deployed HPKP. Well-known for taking security seriously, it sets a plethora of well-configured best-practice security headers.

One of the headers that is set when visiting is the following HPKP header:

Public-Key-Pins: max-age=300;

This HPKP policy specifies two pins, and the directive applies to all subdomains.

The first pinned key (identified by the WoiWRyIOVNa9ihaBciRSC7XHjliYS9VwUGOIud4PB18= SHA-256 hash) corresponds to the DigiCert High Assurance EV Root CA. This is the root of the chain of trust currently used by

The second hash (JbQbUG5JMJUoI6brnx0x3vZF6jilxsapbXGVfjhN8Fg=) is GitHub's backup pin. This corresponds to the VeriSign Class 3 Public Primary Certification Authority - G5 root. As this key does not appear in GitHub's served certificate chain, it is treated as a backup pin.

When GitHub wants to replace its TLS certificate, the new certificate must be signed by either DigiCert or Symantec – otherwise, none of the key hashes in the new certificate chain would match the existing HPKP policy, and its users would be blocked from accessing the site.

Pinning a pair of root certificate keys is arguably less risky than pinning one of GitHub's own backup keys, but there is a rather large trade-off. With GitHub's current HPKP policy, an attacker can still impersonate the site if he can obtain a fraudulent certificate issued by either DigiCert or Symantec. Conversely, if GitHub were to rely on backup keys that only it controlled, then the only way an attacker could impersonate the site is by compromising GitHub's private keys.

Even so, GitHub evidently remains wary — its HPKP header sets a max-age value of 300. This instructs browsers to remember the policy for no longer than 300 seconds, so in the event of a mistake, users will only be denied access for at most five minutes. However, this makes the policy practically toothless.

In the event of an attack, anybody who has not visited the real within the past five minutes is a potential victim. Even if a user has visited GitHub within the past five minutes, being denied access might just be put down to a temporary glitch. A savvy attacker may decide to wait until five minutes after the users last access to GitHub to ensure he will not be caught.


Mozilla is using HPKP much more effectively on its support site, as this site sets a much longer max-age attribute:

Public-Key-Pins: max-age=1296000;

This equates to 15 days, which means it will provide effective protection to anyone who visits the site at least once a fortnight.

Rather than using public key hashes that correspond to more than one certificate authority, Mozilla has chosen to pin to a single CA: both keys are controlled by DigiCert. In some respects, this is a safer policy by ensuring that only a single CA is able to issue new certificates; however, it leaves Mozilla beholden to DigiCert. If DigiCert were ever forced to stop issuance and Mozilla's certificate required replacement, visitors could be locked out of Mozilla's site for up to 15 days.


A much bolder implementation has been deployed by the Pixabay image library on Its Public-Key-Pins header specifies a max-age of one year.


Rather than pinning CA-controlled keys, Pixabay has pinned its own certificate's key, as well as a backup key held by Pixabay. This option trades complete defence against CA compromise with a significant risk if the backup pin cannot be used.

If Pixabay were to lose the private keys for both of these certificates, it would likely be catastrophic – visitors would be denied access to its site for an entire year. Pixabay has evidently decided that robust prevention of impersonation attacks is worth the risk.

Why are so few sites daring to use HPKP?

Only 0.09% of all certificates in Netcraft's March 2016 SSL Survey are using HPKP – that's fewer than 4,100 certificates in the whole world that are being delivered with the Public-Key-Pins header.

If that amount did not already seem astonishingly low, more than a quarter of these sites are using the HPKP header incorrectly, which effectively disables HPKP. Consequently, the total number of certificates that are really using HPKP is actually less than 3,000.

Still in its infancy

One of the reasons why HPKP is so rarely deployed could be that it is a relatively new standard, and is still not supported by all mainstream browsers. However, this only partly explains its poor uptake on servers. Although the Public Key Pinning Extension for HTTP was not formalised until the publication of RFC 7469 in April 2015, a significant proportion of internet users have already been able to benefit from this feature since October 2014, when HPKP support was introduced to Chrome 38.

By the time HPKP support was also added to Firefox 35 in January 2015, around a quarter of all internet users were in a position to benefit from sites using HPKP. But today, HPKP remains unsupported in Internet Explorer, Edge, Safari and Opera Mini. Nonetheless, there are millions of people using browsers that do support HPKP, and the only reason they are not benefiting from this technology is because so few websites are deploying it.

Lack of awareness

Possibly the largest reason for the lack of HPKP deployment is that many website owners are simply unaware that this security feature exists, or do not realise the benefits it can bring. However, this is not the most significant problem for lots of websites, as most also lack simpler features that are widely supported, such as HSTS and "Secure" cookies. Implementing HPKP is largely redundant if a site does not also implement HSTS, as this would still allow a man-in-the-middle attacker to hijack unencrypted HTTP traffic and prevent the victim's browser being redirected to the HTTPS site.

Lack of understanding

Netcraft's SSL Survey shows that lots of trivial mistakes are being made when website administrators try to deploy HPKP headers, which indicates a widespread lack of understanding. The net result of these mistakes is that HPKP is not enabled on many sites.

Fear of the "HPKP Footgun"

HPKP is the best way of protecting a site from being impersonated by mis-issued certificates, but as we have already discussed, it is very easy for this protection to backfire with severe consequences. A small misconfiguration could result in a website becoming inaccessible to its own customers.


HPKP offers a very strong defence against man-in-the-middle attacks, providing it is used in conjunction with HTTPS, HSTS and HSTS Preloading – but despite the obvious security benefits, hardly anyone is using it. Currently, only 0.09% of all secure websites are making use of HPKP headers.

The risk of something going wrong when deploying HPKP is hard to overlook, as a small mistake could ultimately destroy a company's business by making its website inaccessible for months. Only a few thousand secure websites have accepted this risk so far, although you could argue that it only makes sense to deploy HPKP on the largest and most visible websites. For smaller websites, the high risk of something going wrong is outweighed by the incredibly low risk of being attacked: Fraudulently issued certificates are a very rare occurrence, and are more likely to be used to impersonate the biggest websites.

An even newer technology known as Expect CT could potentially provide a safer and easier approach to tackling fraudulently issued certificates. Opted-in websites will be able to tell browsers to expect to see their legitimate certificates in a Certificate Transparency log. These logs are open to public scrutiny, allowing mis-issued certificates to be identified by domain owners, CAs and domain users; and fraudulently issued certificates that do not appear in logs would not be trusted under Expect CT. CAs would be responsible for entering correct details into these logs, thus removing the burden from website operators.

Sites that have properly configured HPKP would be extremely hard to attack in practice, although it is still not impossible. Browsers that have never visited a site before could still be vulnerable to man-in-the-middle attacks if an attacker obtains a valid certificate, because unlike with HSTS, there is no common preload list available for HPKP (it is, however, possible to request special treatment in Google Chrome).

Netcraft offers a range of services that can be used to detect and defeat large-scale pharming attacks, and security testing services that identify man-in-the-middle vulnerabilities in web application and mobile apps. Contact for more information.

March 2016 Web Server Survey

In the March 2016 survey we received responses from 1,003,887,790 sites and 5,782,080 web-facing computers. This reflects a gain of nearly 70 million sites, but a loss of 14,100 computers.

This is the second time the total number of sites has reached more than a billion. This milestone was first reached in September 2014, although it was short-lived: By November 2014, the total fell back below one billion, and had stayed that way until the current month. During the intervening period, the total fell as low as 849 million sites in April 2015.

The total number of websites is typically prone to large fluctuations. Domain holding companies, typo squatters, spammers and link farmers can cause millions of sites to be deployed in a short space of time, without any significant outlay, but these types of site are intrinsically uninteresting to humans. Netcraft's active sites metric counters the effect of these by discounting sites that appear to be automatically generated. This leads to a more-stable metric that better illustrates real, practical use of the web.

The number of active sites currently stands at just 171 million, meaning around 1 in 6 sites are active. The total fell by 764,000 this month, but nginx stands out as being the only major vendor to increase its active site count — by an impressive 699,000. This has increased its active sites share to 16.4%, while Apache's loss of nearly a million active sites took its leading share down to 49.2%.

Typifying nginx's rise amongst active sites, it also showed the only growth in web-facing computers amongst the major server vendors. This month's survey found more than 15,000 additional computers running nginx on the web, while Microsoft's loss of 30,000 computers was the primary cause of the overall loss in this metric. Thankfully, the majority of this decline consisted of Windows Server 2003 computers, which arguably helps improve the safety of the internet — this server software is no longer supported by Microsoft.

China accounts for over 30% of all web-facing computers that run Windows Server 2003, making it the largest user of this obsolete operating system; however, more than half of this month's Windows Server 2003 losses were seen in China, which has helped to bring this share down slightly.

Apache's computer growth was relatively modest at only 447 computers, but Microsoft's large loss caused Apache's market share to increase by 0.12 to 47.9%. nginx's gain of 15,000 computers took its market share up by 0.30 to 14.3%, but Microsoft remains a fair way ahead of nginx with a 26.6% share of the market.

Total number of websites

Web server market share

DeveloperFebruary 2016PercentMarch 2016PercentChange
Continue reading

95% of HTTPS servers vulnerable to trivial MITM attacks

Only 1 in 20 HTTPS servers correctly implements HTTP Strict Transport Security, a widely-supported security feature that prevents visitors making unencrypted HTTP connections to a server.

The remaining 95% are therefore vulnerable to trivial connection hijacking attacks, which can be exploited to carry out effective phishing, pharming and man-in-the-middle attacks. An attacker can exploit these vulnerabilities whenever a user inadvertently tries to access a secure site via HTTP, and so the attacker does not even need to spoof a valid TLS certificate. Because no crypto-wizardry is required to hijack an HTTP connection, these attacks are far easier to carry out than those that target TLS, such as the recently announced DROWN attack.


The growth of HTTPS has been a mostly positive step in the evolution of the internet, enabling encrypted communications between more users and websites than ever before. Many high profile sites now use HTTPS by default, and millions of TLS certificates are currently in use on the web. With companies like Let's Encrypt offering free certificates and automated management tools, it is also easier than ever to deploy an HTTPS website that will be trusted by all modern browsers.

The primary purpose of a TLS certificate is to allow a browser to verify that it is communicating with the correct website. For example, if uses a valid TLS certificate, then a man-in-the-middle attacker would not be able to hijack a browser's connection to this site unless he is also able to obtain a valid certificate for that domain.

A man-in-the-middle attack like this is generally not possible if the customer uses HTTPS.

A man-in-the-middle attack like this is generally not possible if the initial request from the customer uses HTTPS.

It would be extremely difficult for the attacker to obtain a valid certificate for a domain he does not control, and using an invalid certificate would cause the victim's browser to display an appropriate warning message. Consequently, man-in-the-middle attacks against HTTPS services are hard to pull off, and often not very successful. However, there are plenty of realistic opportunities to use the unencrypted HTTP protocol to attack most HTTPS websites.

HTTP Strict Transport Security (HSTS)

Encrypted communications are an essential requirement for banks and other financial websites, but HTTPS alone is not sufficient to defend these sites against man-in-the-middle attacks. Astonishingly, many banking websites lurk amongst the 95% of HTTPS servers that lack a simple feature that renders them still vulnerable to pharming and man-in-the-middle attacks. This missing feature is HTTP Strict Transport Security (HSTS), and only 1 in 20 secure servers currently make use of it, even though it is supported by practically all modern browsers.

Each secure website that does not implement an HSTS policy can be attacked simply by hijacking an HTTP connection that is destined for it. This is a surprisingly feasible attack vector, as there are many ways in which a user can inadvertently end up connecting via HTTP instead of HTTPS.

Manually typed URLs often result in an initial insecure request, as most users do not explicitly type in the protocol string (http:// or https://). When no protocol is given, the browser will default to HTTP – unless there is an appropriate HSTS policy in force.

To improve accessibility, most secure websites also run an HTTP service to redirect users to the corresponding HTTPS site – but this makes them particularly prone to man-in-the-middle attacks if there is no HSTS policy in force. Not only would many users be accustomed to visiting the HTTP site first, but anyone else who visits the site via an old bookmark or search engine result might also initially access the site via an insecure HTTP address. Whenever this happens, the attacker can hijack the initial HTTP request and prevent the customer being redirected to the secure HTTPS website.

This type of attack can be automated with the sslstrip tool, which transparently hijacks HTTP traffic on a network and converts HTTPS links and redirects into HTTP. This type of exploit is sometimes regarded as a protocol downgrade attack, but strictly speaking, it is not: rather than downgrading the protocol, it simply prevents the HTTP protocol being upgraded to HTTPS.

NatWest's online banking website at lacks an HSTS policy and also offers an HTTP service to redirect its customers to the HTTPS site. This setup is vulnerable to the type of man-in-the-middle attack described above.

NatWest's online banking website at lacks an HSTS policy and also offers an HTTP service to redirect its customers to the HTTPS site. This setup is vulnerable to the type of man-in-the-middle attack described above.

Vulnerable sites can be attacked on a massive scale by compromising home routers or DNS servers to point the target hostname at a server that is controlled by the attacker (a so-called "pharming" attack). Some smaller scale attacks can be carried out very easily – for example, if an attacker sets up a rogue Wi-Fi access point to provide internet access to nearby victims, he can easily influence the results of their DNS lookups.

Even if a secure website uses HTTPS exclusively (i.e. with no HTTP service at all), then man-in-the-middle attacks are still possible. For example, if a victim manually types into his browser's address bar—without prefixing it with https://—the browser will attempt to make an unencrypted HTTP connection to, even if the genuine site does not run an HTTP service. If this hostname has been pharmed, or is otherwise subjected to a man-in-the-middle attack, the attacker can hijack the request nonetheless and eavesdrop the connection as it is relayed to the genuine secure site, or serve phishing content directly to the victim.

In short, failing to implement an HSTS policy on a secure website means attackers can carry out man-in-the-middle attacks without having to obtain a valid TLS certificate. Many victims would fall for these attacks, as they can be executed over an unencrypted HTTP connection, thus avoiding any of the browser's tell-tale warnings about invalid certificates.

Implementing HSTS: A simple one-liner

The trivial man-in-the-middle attacks described above can be thwarted by implementing an appropriate HSTS policy. A secure website can do this simply by setting a single HTTP header in its responses:

    Strict-Transport-Security: max-age=31536000;

This header can only be set over an HTTPS connection, and instructs compatible browsers to only access the site over HTTPS for the next year (31,536,000 seconds = 1 year). This is the most common max-age value, used by nearly half of all HTTPS servers. After this HSTS policy has been applied, even if a user manually prefixes the site's hostname with http://, the browser will ignore this and access the site over HTTPS instead.

The combination of HSTS and HTTPS therefore provides a good defence against pharming attacks, as the attacker will not be able to redirect and intercept plaintext HTTP traffic when a client obeys the HSTS policy, nor will he be able to present a valid TLS certificate for the site he is impersonating.

The attacker cannot even rely on a small proportion his victims unwisely ignoring the use of an invalid certificate, as browsers must regard this situation as a hard fail when an HSTS policy is in force. The browser will simply not let the victim access the site if it finds an invalid certificate, nor will it allow an exception to be added.

When Google Chrome encounters an invalid certificate for a site that has an effective HSTS policy, the victim is not allowed to bypass the browser's warning message or add an exception.

When Google Chrome encounters an invalid certificate for a site that has an effective HSTS policy, the victim is not allowed to bypass the browser's warning message or add an exception.

To prevent other types of attack, it is also wise to add the includeSubDomains directive to ensure that every possible subdomain of a site is protected by HSTS. This mitigates cookie injection and session fixation attacks that could be executed by impersonating an HTTP site on a non-existent subdomain such as, and using it to set a cookie which would be sent to the secure site at This directive can be enabled like so:

    Strict-Transport-Security: max-age=31536000; includeSubDomains

However, some thought is required before taking the carte blanche approach of including all subdomains in an HSTS policy. The website's administrators must ensure that every single one of its subdomains supports HTTPS for at least the duration specified by the max-age parameter, otherwise users of these subdomains risk being locked out.

Setting an HSTS policy will also protect first time visitors who habitually use search bars or search engines to reach their destination. For example, typing "paypal" into Google's HTTPS search engine will yield a link to, because Google will always link to the HTTPS version of a website if an appropriate HSTS policy exists.

HSTS preloading

HSTS is clearly an important security feature, but there are several circumstances under which its benefits will not work. Because HSTS directives are delivered via an HTTP header (over an HTTPS connection), HSTS can only instruct a browser to only use HTTPS after the browser's first visit to a secure website.

Men-in-the-middle can therefore still carry out attacks against users who have:

  • Never before visited the site.
  • Recently reinstalled their operating system.
  • Recently reinstalled their browser.
  • Switched to a new browser.
  • Switched to a new device (e.g. mobile phone).
  • Deleted their browser's cache.
  • Not visited the site within the past year (or however long the max-age period lasts).

These vulnerabilities can be eliminated by using HSTS Preloading, which ensures that the site's HSTS policy is distributed to supported browsers before the customer's first visit.

Website administrators can use the form at to request for domains to be included in the HSTS Preload list maintained by Google. Each site must have a valid certificate, redirect all HTTP traffic to HTTPS, and serve all subdomains over HTTPS. The HSTS header served from each site must specify a max-age of at least 18 weeks (10,886,400 seconds) and include the preload and includeSubdomains directives.

It can take several months for domains to be reviewed and propagated to the latest stable versions of Firefox, Safari, Internet Explorer, Edge and Chrome. When domains are added to the preload list, all users of these browsers will benefit from the security offered by HSTS, even if they have never visited the sites before.


HSTS is widely supported, but not widely implemented. Nearly all modern browsers obey HSTS policies, including Internet Explorer 11, Microsoft Edge, Firefox, Chrome, Safari and Opera – yet less than 5% of secure websites enable this important security feature.

Secure websites that do not use HSTS are trivial to attack if the attacker can hijack a victim's web traffic, but it is even easier to defeat such attacks by implementing an HSTS policy. This begs the question of why so few websites are using HSTS.

The HSTS specification (RFC 6797) was published in 2012, and so it can hardly be considered a new technology any more. Nonetheless, many website administrators might still be unaware of its existence, or may not yet feel ready to commit to running an HTTPS-only website. These are probably the most significant reasons for its low uptake.

Some website administrators have even disabled HSTS by explicitly setting a max-age of 0 seconds. This has the effect of switching off any previously established HSTS policies, but this backpedalling can only take proper effect if every client revisits the secure site after the max-age has been set to zero. When a site implements an HSTS policy, it is effectively committed to maintaining its HTTPS service for as long as the largest max-age it has ever specified, otherwise it risks denying access to infrequent visitors. Nearly 4% of all HTTPS servers that use the Strict-Transport-Security header currently set a max-age of zero, including Twitter's URL-shortener.

Browser support for HSTS can also introduce some privacy concerns. By initiating requests to several distinct hostnames (some of which enable HSTS), a hostile webpage can establish a "supercookie" to uniquely identify the client browser during subsequent visits, even if the user deletes the browser's conventional cookies. The browser will remember which pattern of hostnames had HSTS enabled, thus allowing the supercookie to persist. However, this privacy concern only affects clients and does not serve as an excuse for websites to avoid implementing their own HSTS policies.

Implementing an HSTS policy is very simple and there are no practical downsides when a site already operates entirely over HTTPS. This makes it even more surprising to see many banks failing to use HSTS, especially on their online banking platforms. This demonstrates poor security practices where it matters the most, as these are likely to be primary targets of pharming attacks.

Netcraft offers a range of services that can be used to detect and defeat large-scale pharming attacks, and security testing services that identify man-in-the-middle vulnerabilities in web application and mobile apps. Contact for more information.

Most Reliable Hosting Company Sites in February 2016

Rank Performance Graph OS Outage
DNS Connect First
1 Inc Linux 0:00:00 0.000 0.266 0.005 0.015 0.016
2 Netcetera Linux 0:00:00 0.004 0.077 0.087 0.178 0.178
3 Lightcrest unknown 0:00:00 0.009 0.301 0.006 0.022 0.026
4 XILO Communications Ltd. Linux 0:00:00 0.009 0.235 0.075 0.143 0.143
5 Bigstep Linux 0:00:00 0.013 0.151 0.062 0.124 0.124
6 Qube Managed Services Linux 0:00:00 0.018 0.152 0.060 0.128 0.128
7 Linux 0:00:00 0.018 0.485 0.203 0.422 0.422
8 Datapipe Linux 0:00:00 0.022 0.156 0.013 0.026 0.032
9 LeaseWeb Linux 0:00:00 0.022 0.368 0.026 0.054 0.054
10 Anexia Linux 0:00:00 0.022 0.194 0.086 0.177 0.177

See full table

GoDaddy had the most reliable hosting company site in February, responding to every Netcraft request. This is the ninth consecutive month that GoDaddy has featured in the top 10, and the fourth time it has reached first place in the past 12 months.

Netcetera took second place in February with just one failed request. Based in the Isle of Man, Netcetera is celebrating its 20th birthday in 2016 with a series of special promotions. Over 11 years of monitoring by Netcraft, Netcetera has maintained a 99.96% uptime record.

Third place goes to Lightcrest, appearing in the top 10 for the third time in four months. Although both Lightcrest and XILO had two failed requests each, Lightcrest had a faster average connection time. Lightcrest's customers can use its Kahu Compute Fabric platform for Lightcrest-hosted deployments and those hosted remotely.

Linux remains the dominant choice of operating system among the most reliable hosting sites. This is the eighth consecutive month without any sites powered by Windows appearing in the top 10.

Netcraft measures and makes available the response times of around forty leading hosting providers' sites. The performance measurements are made at fifteen minute intervals from separate points around the internet, and averages are calculated over the immediately preceding 24 hour period.

From a customer's point of view, the percentage of failed requests is more pertinent than outages on hosting companies' own sites, as this gives a pointer to reliability of routing, and this is why we choose to rank our table by fewest failed requests, rather than shortest periods of outage. In the event the number of failed requests are equal then sites are ranked by average connection times.

Information on the measurement process and current measurements is available.

February 2016 Web Server Survey

In the February 2016 survey we received responses from 933,892,520 sites and 5,796,210 web-facing computers.

Microsoft has edged closer towards Apache, with an increase of 16.1 million sites bringing its total up by 6.14% to 279 million. Apache's relatively modest growth of 0.66% has put Microsoft within 3 percentage points of Apache's leading market share of 32.8%.

In terms of web-facing computers, Apache maintains a much clearer lead with a 47.8% share of the market. Microsoft also takes second place by this metric, albeit with a share that is more than 20 percentage points behind Apache. However, both Apache and Microsoft suffered small losses in market share as nginx continues to exhibit strong growth: This month, nginx gained 21,100 computers, increasing its market share by 0.26 points to 13.96%.

nginx 1.9.11 mainline was released on 9th February. This version introduced support for dynamic modules, enabling selective loading of both third-party and some native modules at runtime. In previous versions, nginx modules had to be statically linked into an nginx binary built from source, causing the module to be loaded every time even if it was not going to be used.

The latest version of Microsoft Internet Information Services, IIS 10.0, is still very rare in the wild, as its primary deployment platform (Windows Server 2016) has yet to be released. Fewer than 5,000 websites are currently using IIS 10.0, and these are being served either from technical preview versions of Windows Server 2016, or from Windows 10 machines.

The latest technical preview version of Windows Server 2016 also supports a headless deployment option known as Nano Server. This is a stripped-down version of Windows Server, without a graphical interface and a few other features that are not essential for modern web applications. As a result, it typically requires fewer updates to be installed – and consequently, fewer reboots, too.

Despite losing a small amount of market share, Apache also showed a reasonable growth of 15,600 computers. Similar to last month, a significant proportion of this growth was due to the appearance of more Western Digital My Cloud consumer storage devices.

The total number of My Cloud devices in the survey now stands at 583,400, which is 68,400 more than last month; however, the number of devices that are exposed directly to the internet grew by only 11,100.

Western Digital is using Amazon AWS to host the servers that proxy requests to My Cloud devices in Relay mode. Most of these relay servers have been configured to serve a few thousand devices each, and so the 331,000 devices that are currently using Relay mode contribute fewer than 200 computers towards Apache's total.

Interestingly, while most web-facing My Cloud devices are hosted in the US, more than half of the * hostnames used by the relay servers are hosted in Amazon's EU regions.

Total number of websites

Web server market share

DeveloperJanuary 2016PercentFebruary 2016PercentChange
Continue reading

eBay scripting flaws being actively exploited by fraudsters

Fraudsters are actively exploiting scripting flaws in eBay's auction platform to carry out a spate of highly convincing scams. The latest attacks steal money and credentials, and are still taking place today, despite a recent partial fix that attempted to stop these flaws being exploited.

Victims stand to lose thousands of pounds in these attacks.

Victims stand to lose thousands of pounds in these clever phishing attacks, which are launched directly from eBay's own website.

eBay had originally declined to fix the vulnerability at all, resulting in widespread criticism from security experts, which perhaps influenced its subsequent decision to implement a partial fix. eBay later said that it took the issue very seriously, and that it did not find any fraudulent activity stemming from this incident.

However, this week, Netcraft has seen several fraudulent eBay listings that actively exploit these flaws. Not only does this demonstrate that the underlying issue has not been adequately fixed, but it shows that it is also being exploited by fraudsters.

Despite the partial fix, fraudsters are still able to include malicious JavaScript in eBay listing descriptions, and these scripts are being used to great effect. Merely viewing a fraudster's auction will cause a user to be automatically diverted away from the genuine eBay website and onto a phishing site.

These latest incidents rank amongst the most impressive phishing attacks, as they are so incredibly convincing. They are launched from the genuine eBay website, which will have already gained the victim's trust, and no untoward user interaction is required. Because the phishing site looks so similar to the genuine eBay website, it is likely that most victims would never realise they have ended up on a fraudulent website, especially as they did not click on any links pointing to external sites.

This week's attacks appeared to be posted from compromised eBay accounts, some of which had been created several years ago. This makes it difficult for victims to identify the listings as fraudulent, as they are ostensibly posted by users who have been members for several years and have 100% positive feedback.

A real attack in detail

Motor vehicles are a magnet for fraudulent activity on eBay due to the high values of such items. Fraudsters will often copy the contents of a previous, legitimate listing and use it to create their own listing at a temptingly low price.

Three months ago, the following motorhome was sold on eBay for £19,295. This was a genuine listing for a genuine vehicle, so some details have been redacted to protect the innocent:

A legitimate eBay auction that ended three months ago. The details of this motorhome auction were reused in one of this week's fraudulent listings.

A legitimate eBay auction that ended three months ago. The details of this motorhome auction were reused in one of this week's fraudulent listings.

This week, a fraudster copied the contents of this auction and used it to create his own fraudulent listing. This would likely be found by anyone searching for similar motorhomes:

The fraudulent listing, as it appeared in eBay's search results.

The fraudulent listing, as it appeared in eBay's search results.

The fraudulent listing was posted from an eBay account created in April 2010. The user has 100% positive feedback accrued from previous, legitimate purchases, which suggests that the account had been compromised by the fraudster. The fraudulent listing has since been removed from eBay, but the legitimate account remains.

The compromised account that was used to post the fraudulent listing had 100% positive feedback.

The compromised account that was used to post the fraudulent listing had 100% positive feedback.

When a victim views the fraudulent listing, he is immediately redirected to the fraudster's phishing site on Other than the domain name, this looks practically identical to the genuine eBay website, and also features the same username and feedback score from the compromised seller's account:

The victim is automatically redirected to this phishing site when he views the fraudulent listing on eBay. This site was reported to Netcraft and blocked.

The victim is automatically redirected to this phishing site when he views the fraudulent listing on eBay. This site was reported to Netcraft and blocked.

The significantly discounted "Buy it now" price of £6,300 would be extremely alluring to a prospective buyer, considering that the real thing sold for nearly £20,000 in a legitimate auction three months ago.

The redirection to the phishing site is carried out automatically as soon as the victim views the fraudulent listing on the eBay website. The item's description is copied from the legitimate listing we saw three months ago, but a malicious block of JavaScript has been added.

The malicious script appended to the description of the motorhome. Some of this has been blurred to prevent copycat attacks until eBay properly fixes the vulnerability.

The malicious script appended to the description of the motorhome. We have blurred the salient parts of this code to prevent copycat attacks until eBay properly fixes the vulnerability.

The malicious script has been specially constructed in order to bypass the cross-site scripting filters that eBay has implemented.

eBay disallows certain strings in the item description when a listing is created, but this week's attacks demonstrate that this security measure is insufficient.

eBay disallows certain strings in the item description when a listing is created, but this week's attacks demonstrate that this security measure is still insufficient.

When the external script is executed by the victim's browser, it redirects the victim to a URL redirection service, TinyURL, which in turn redirects him to the fraudster's phishing site.

The externally hosted JavaScript, which is executed by the fraudulent eBay listing. This site uses a .space domain that was registered on 6 February.

The externally hosted JavaScript, which is executed by the fraudulent eBay listing. This file is served from a .space domain that was registered on 6 February.

After the above JavaScript has redirected the victim to the phishing site, a server-side PHP script named php.php uses a random number generator to create a new PHP script, such as 54388632.php. The victim is then redirected to the newly-created script, which displays the fraudulent content. The randomly-named file is then deleted from the web server. If the victim were to notice that he was on a phishing site, it would be difficult to report it to anyone, as the URL in the victim's address bar would lead to a 404 Not Found error page on any subsequent visit.

The phishing site uses a domain name that was registered through less than two weeks ago, and there is no content on the site's homepage. This indicates that the domain was probably registered specifically for use in these fraudulent eBay listings. Both this site and the malicious JavaScript file on are hosted by HostGator.

If the victim attempts to ask the seller a question, he will be taken to an enquiry form that is hosted on the fraudster's phishing site. Any questions asked here are sent directly to the fraudster.

Any questions asked about the item are sent directly to the fraudster.

Any questions asked about the item are sent directly to the fraudster.

If the victim tries to make a Best Offer for the vehicle, he is prompted to enter an email address. This address is later used by the fraudster to solicit payment directly from the victim, often via bank transfer.

After making an offer, the victim can soon expect to hear from the fraudster.

After making an offer, the victim can soon expect to hear from the fraudster.

This particular phishing attack demonstrates some interesting evolutions in the fraudsters' methodologies. Not only is it rather cleverly launched from the legitimate eBay site, and uses randomly-named files that are deleted to evade detection, but it also tries to avoid leaving any evidence in eBay's server logs: While all of the pictures used on the spoof auction page are stolen from the earlier legitimate auction, they are either encoded as inline Base64-encoded images, or are served from the fraudster's own website. This means that no Referer headers will be transmitted to eBay's web servers, which would otherwise give away the location of the phishing site.

This phishing attack is unusual in that it does not attempt to steal the victim's eBay password or any other account credentials. This subtlety could contribute to its effectiveness, as some victims might more readily identify a scam that does ask for a password.

The victim's offer and email address is all the fraudster needs in order to solicit payment. To instil further trust in the victim, these payment requests usually claim to use a third-party escrow service to accept the money. A genuine escrow service would release the money to the seller only if the customer receives the goods they paid for, but unsurprisingly, these eBay vehicle scams do not use a real escrow service. When the victim transfers his money to the specified account, it goes straight to the fraudster.

A fake escrow email from a fraudulent car seller. This one purportedly related to the sale of a Volkswagen T5 Transporter.

A fake escrow email from a fraudulent car seller. This one purportedly related to the sale of a Volkswagen T5 Transporter.

To discourage the victim from visiting their bank, who might warn him that it is a scam, the email adds: "You can pay using your online baking [sic], because it saves a considerable amount of time. Online Banking saves you the trouble of going to a bank and wasting your valuable time (payments can also be made on the weekends)."

Old habits die hard

Netcraft highlighted the risks posed by allowing JavaScript in eBay listings almost two years ago, when a series of similar attacks took place. eBay's only apparent protection against these attacks was a policy that we demonstrated can be easily ignored by fraudsters.

As eBay's latest fix is only a "partial" one, it suggests that eBay still might not have any intention of completely fixing these vulnerabilities. eBay previously explained that allowing active content in legitimate listings is worth the security risk, as the benefits outweigh the likelihood of being attacked.

A plea for help: This fraud victim claims to have been scammed on eBay after sending a bank transfer to pay for a caravan.

A plea for help: This fraud victim claims to have been scammed on eBay this week after sending a bank transfer to pay for a caravan.

These attacks have continued throughout the week. The following example was found earlier today – victims were redirected to this phishing site after viewing yet another specially crafted listing on the real eBay website.


Another eBay phishing site, which victims are automatically redirected to after viewing a specially crafted listing on the real eBay website.

It is likely that both examples have been orchestrated by the same fraudster, as both domain names were registered through the same company two weeks ago. However, today's example also attempts to steal the victim's eBay username and password when the victim clicks the Buy it now button.

The latest example also tries to steal the victim's eBay username and password.

The latest example also tries to steal the victim's eBay username and password.

The fraudster can use these stolen credentials to create additional fraudulent listings on his victims' own eBay accounts, which in turn can be used to steal more accounts and more money. This is a cycle of fraud that will be difficult to stop if eBay does not fully resolve this vulnerability.