Cached and Confused: Web Cache Deception in the Wild

12/21/2019 ∙ by Seyed Ali Mirheidari, et al. ∙ 0

Web cache deception (WCD) is an attack proposed in 2017, where an attacker tricks a caching proxy into erroneously storing private information transmitted over the Internet and subsequently gains unauthorized access to that cached data. Due to the widespread use of web caches and, in particular, the use of massive networks of caching proxies deployed by content distribution network (CDN) providers as a critical component of the Internet, WCD puts a substantial population of Internet users at risk. We present the first large-scale study that quantifies the prevalence of WCD in 340 high-profile sites among the Alexa Top 5K. Our analysis reveals WCD vulnerabilities that leak private user data as well as secret authentication and authorization tokens that can be leveraged by an attacker to mount damaging web application attacks. Furthermore, we explore WCD in a scientific framework as an instance of the path confusion class of attacks, and demonstrate that variations on the path confusion technique used make it possible to exploit sites that are otherwise not impacted by the original attack. Our findings show that many popular sites remain vulnerable two years after the public disclosure of WCD. Our empirical experiments with popular CDN providers underline the fact that web caches are not plug play technologies. In order to mitigate WCD, site operators must adopt a holistic view of their web infrastructure and carefully configure cache settings appropriate for their applications.



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Web caches have become an essential component of the Internet infrastructure with numerous use cases such as reducing bandwidth costs in private enterprise networks and accelerating content delivery over the World Wide Web. Today caching is implemented at multiple stages of Internet communications, for instance in popular web browsers [45, 58], at caching proxies [64, 55], and directly at origin web servers [6, 46].

In particular, Content Delivery Network (CDN) providers heavily rely on effective web content caching at their edge servers, which together comprise a massively-distributed Internet overlay network of caching reverse proxies. Popular CDN providers advertise accelerated content delivery and high availability via global coverage and deployments reaching hundreds of thousands of servers [16, 5]

. A recent scientific measurement also estimates that more than 74% of the Alexa Top 1K are served by CDN providers, indicating that CDNs and more generally web caching play a central role in the Internet 


While there exist technologies that enable limited caching of dynamically-generated pages, web caching primarily targets static, publicly accessible content. In other words, web caches store static content that is costly to deliver due to an object’s size or distance. Importantly, these objects must not contain private or otherwise sensitive information, as application-level access control is not enforced at cache servers. Good candidates for caching include frequently accessed images, software and document downloads, streaming media, style sheets, and large static HTML and JavaScript files.

In 2017, Gil presented a novel attack called web cache deception (WCD) that can trick a web cache into incorrectly storing sensitive content, and consequently give an attacker unauthorized access to that content [23, 24]. Gil demonstrated the issue with a real-life attack scenario targeting a high profile site, PayPal, and showed that WCD can successfully leak details of a private payment account. Consequently, WCD garnered significant media attention, and prompted responses from major web cache and CDN providers [43, 12, 9, 8, 48, 13].

At its core, WCD results from path confusion between an origin server and a web cache. In other words, different interpretations of a requested URL at these two points lead to a disagreement on the cacheability of a given object. This disagreement can then be exploited to trick the web cache into storing non-cacheable objects. WCD does not imply that these individual components—the origin server and web cache—are incorrectly configured per se. Instead, their hazardous interactions as a system lead to the vulnerability. As a result, detecting and correcting vulnerable systems is a cumbersome task, and may require careful inspection of the entire caching architecture. Combined with the aforementioned pervasiveness and critical role of web caches in the Internet infrastructure, WCD has become a severely damaging issue.

In this paper, we first present a large-scale measurement and analysis of WCD over 295 sites in the Alexa Top 5K. We present a repeatable and automated methodology to discover vulnerable sites over the Internet, and a detailed analysis of our findings to characterize the extent of the problem. Our results show that many high-profile sites that handle sensitive and private data are impacted by WCD and are vulnerable to practical attacks. We then discuss additional path confusion methods that can maximize the damage potential of WCD, and demonstrate their impact in a follow-up experiment over an extended data set of 340 sites.

To the best of our knowledge, this is the first in-depth investigation of WCD in a scientific framework and at this scale. In addition, the scope of our investigation goes beyond private data leakage to provide novel insights into the severity of WCD. We demonstrate how WCD can be exploited to steal other types of sensitive data including security tokens, explain advanced attack techniques that elevate WCD vulnerabilities to injection vectors, and quantify our findings through further analysis of collected data.

Finally, we perform an empirical analysis of popular CDN providers, documenting their default caching settings and customization mechanisms. Our findings underline the fact that WCD is a system safety problem. Site operators must adopt a holistic view of their infrastructure, and carefully configure web caches taking into consideration their complex interactions with origin servers.

To summarize, we make the following contributions:

  • We propose a novel methodology to detect sites impacted by WCD at scale. Unlike existing WCD scan tools that are designed for site administrators to test their own properties in a controlled environment, our methodology is designed to automatically detect WCD in the wild.

  • We present findings that quantify the prevalence of WCD in 295 sites among the Alexa Top 5K, and provide a detailed breakdown of leaked information types. Our analysis also covers security tokens that can be stolen via WCD as well as novel security implications of the attack, all areas left unexplored by existing WCD literature.

  • We conduct a follow-up measurement over 340 sites among the Alexa Top 5K that show variations on the path confusion technique make it possible to successfully exploit sites that are not impacted by the original attack.

  • We analyze the default settings of popular CDN providers and document their distinct caching behavior, highlighting that mitigating WCD necessitates a comprehensive examination of a website’s infrastructure.

Ethical Considerations. We have designed our measurement methodology to minimize the impact on scanned sites, and limit the inconvenience we impose on site operators. Similarly, we have followed responsible disclosure principles to notify the impacted parties, and limited the information we share in this paper to minimize the risk of any inadvertent damage to them or their end-users. We discuss details of the ethical considerations pertaining to this work in Section 3.5.

2 Background & Related Work

In this section, we present an overview of how web cache deception (WCD) attacks work and discuss related concepts and technologies such as web caches, path confusion, and existing WCD scanners. As of this writing, the academic literature has not yet directly covered WCD. Nevertheless, in this section we summarize previous publications pertaining to other security issues around web caches and CDNs.

2.1 Web Caches

Repeatedly transferring heavily used and large web objects over the Internet is a costly process for both web servers and their end-users. Multiple round-trips between a client and server over long distances, especially in the face of common technical issues with the Internet infrastructure and routing problems, can lead to increased network latency and result in web applications being perceived as unresponsive. Likewise, routinely accessed resources put a heavy load on web servers, wasting valuable computational cycles and network bandwidth. The Internet community has long been aware of these problems, and deeply explored caching strategies and technologies as an effective solution.

Today web caches are ubiquitous, and are used at various—and often multiple—steps of Internet communications. For instance, client applications such as web browsers implement their own private cache for a single user. Otherwise, web caches deployed together with a web server, or as a man-in-the-middle proxy on the communication path implement a shared cache designed to store and serve objects frequently accessed by multiple users. In all cases, a cache hit eliminates the need to request the object from the origin server, improving performance for both the client and server.

In particular, web caches are a key component of Content Delivery Networks (CDN) that provide web performance and availability services to their users. By deploying massively-distributed networks of shared caching proxies (also called edge servers) around the globe, CDNs aim to serve as many requests as possible from their caches deployed closest to clients, offloading the origin servers in the process. As a result of multiple popular CDN providers that cover different market segments ranging from simple personal sites to large enterprises, web caches have become a central component of the Internet infrastructure. A recent study by Guo et al. estimates that 74% of the Alexa Top 1K make use of CDNs [26].

The most common targets for caching are static but frequently accessed resources. These include static HTML pages, scripts and style sheets, images and other media files, and large document and software downloads. Due to the shared nature of most web caches, objects containing dynamic, personalized, private, or otherwise sensitive content are not suitable for caching. We point out that there exist technologies such as Edge Side Includes [63] that allow caching proxies to assemble responses from a cached static part and a freshly-retrieved dynamic part, and the research community has also explored caching strategies for dynamic content. That being said, caching of non-static objects is not common, and is not relevant to WCD attacks. Therefore, it will not be discussed further in this paper.

The HTTP/1.1 specification defines Cache-Control headers that can be included in a server’s response to signal to all web caches on the communication path how to process the transferred objects [21]. For example, the header ‘‘Cache- Control: no-store’’ indicates that the response should not be stored. While the specification states that web caches MUST respect these headers, web cache technologies and CDN providers offer configuration options for their users to ignore and override header instructions. Indeed, a common and easy configuration approach is to create simple caching rules based on resource paths and file names, for instance, instructing the web cache to store all files with extensions such as jpg, ico, css, or js [18, 15].

2.2 Path Confusion

Traditionally, URLs referenced web resources by directly mapping these to a web server’s filesystem structure, followed by a list of query parameters. For instance, would correspond to the file home/index.html at that web server’s document root directory, and lang=en represents a parameter indicating the preferred language.

However, as web applications grew in size and complexity, web servers introduced sophisticated URL rewriting mechanisms to implement advanced application routing structures as well as to improve usability and accessibility. In other words, web servers parse, process, and interpret URLs in ways that are not clearly reflected in the externally-visible representation of the URL string. Consequently, the rest of the communication endpoints and man-in-the-middle entities may remain oblivious to this additional layer of abstraction between the resource filesystem path and its URL, and process the URL in an unexpected—and potentially unsafe—manner. This is called path confusion.

The widespread use of clean URLs (also known as RESTful URLs) help illustrate this disconnect and the subsequent issues resulting from different interpretations of a URL. Clean URL schemes use structures that abstract away from a web server’s internal organization of resources, and instead provide a more readable API-oriented representation. For example, a given web service may choose to structure the URL as in clean URL representation. Now, consider the case where a user accesses the same web service using the URL The user and all technologies in the communication path (e.g., the web browser, caches, proxies, web application firewalls) are likely to misinterpret this request, expect an image file in return, and treat the HTTP response accordingly (e.g., web caches may choose to store the response payload). However, in reality, the web service will internally map this URL to, and return the contents of index.php with an HTTP 200 status code. Note that even when img/pic.jpg is an arbitrary resource that does not exist on the web server, the HTTP 200 status code will falsely indicate that the request was successfully handled as intended.

Web application attacks that involve malicious payload injection, such as cross-site scripting, are well-understood and studied by both academics and the general security community. Unfortunately, the security implications of path confusion have started to garner attention only recently, and academic literature on the subject is sparse.

One notable class of attacks based on path confusion is Relative Path Overwrite (RPO), first presented by Gareth Heyes in 2014 [28]. RPO targets sites that utilize relative paths for security-sensitive resource inclusions such as style sheets and scripts. The attack is made possible by maliciously-crafted URLs that are still interpreted in the same way their benign counterparts are by web servers, but when used as the base URL causes a web browser to expand relative paths incorrectly. This results in attacker-controlled same-origin inclusions. Other researchers have since proposed variations on more advanced applications of RPO, which can elevate this attack vector into numerous other vulnerabilities [33, 17, 36, 57]. Recently, Arshad et al. conducted a large-scale measurement study of RPO in the wild and reported that 9% of the Alexa Top 1M are vulnerable, and that more than one third of these are exploitable [7].

Other related work include more general techniques for exploiting URL parser behavior. For instance, Orange Tsai presented a series of exploitation techniques that take advantage of the quirks of built-in URL parsers in popular programming languages and web frameworks [61, 62]. While Tsai’s discussion mainly focuses on Server-Side Request Forgery, these techniques are essentially instances of path confusion and can be utilized in many attacks in the category.

Our focus in this paper is web cache deception, the most recently discovered major security issue that is enabled by an attacker exploiting a path confusion vulnerability. To the best of our knowledge, this paper is the first academic exploration of WCD in the literature, and also constitutes the first large-scale analysis of its spread and severity.

Figure 1: An illustrated example of web cache deception. Path confusion between a web cache and a web server leads to unexpected caching of the victim’s private account details. The attacker can then issue a request resulting in a cache hit, gaining unauthorized access to cached private information.

2.3 Web Cache Deception

WCD is a recently-discovered manifestation of path confusion that an attacker can exploit to break the confidentiality properties of a web application. This may result in unauthorized disclosure of private data belonging to end-users of the target application, or give the attacker access to sensitive security tokens (e.g., CSRF tokens) that could be used to facilitate further web application attacks by compromising authentication and authorization mechanisms. Gil proposed WCD in 2017, and demonstrated its impact with a practical attack against a major online payment provider, PayPal [23, 24].

In order to exploit a WCD vulnerability, the attacker crafts a URL that satisfies two properties:

  1. The URL must be interpreted by the web server as a request for a non-cacheable page with private information, and it should trigger a successful response.

  2. The same URL must be interpreted by an intermediate web cache as a request for a static object matching the caching rules in effect.

Next, the attacker uses social engineering channels to lure a victim into visiting this URL, which would result in the incorrect caching of the victim’s private information. The attacker would then repeat the request and gain access to the cached contents. Figure 1 illustrates these interactions.

In Step 1 , the attacker tricks the victim into visiting a URL that requests /account.php/nonexistent.jpg. At a first glance this appears to reference an image file, but in fact does not point to a valid resource on the server. In Step 2 , the request reaches the web server and is processed. The server in this example applies rewrite rules to discard the non-existent part of the requested object, a common default behavior for popular web servers and application frameworks. As a result, the server sends back a success response, but actually includes the contents of account.php in the body, which contains private details of the victim’s account. Unaware of the URL mapping that happened at the server, the web cache stores the response, interpreting it as a static image. Finally, in Step 3 , the attacker visits the same URL which results in a cache hit and grants him unauthorized access to the victim’s cached account information.

Using references to non-existent cacheable file names that are interpreted as path parameters is an easy and effective path confusion technique to mount a WCD attack, and is the original attack vector proposed by Gil. However, we discuss novel and more advanced path confusion strategies in Section 5. Also note that the presence of a Cache-Control: no-store header value has no impact in our example, as it is common practice to enable caching rules on proxy services that simply ignore header instructions and implement aggressive rules based on path and file extension patterns (see Section 2.1).

WCD garnered significant media attention due to its security implications and high damage potential. Major web cache technology and CDN providers also responded, and some published configuration hardening guidelines for their customers [43, 9, 8]. More recently, Cloudflare announced options for new checks on HTTP response content types to mitigate the attack [12].

Researchers have also published tools to scan for and detect WCD, for instance, as an extension to the Burp Suite scanner or as stand-alone tools [31, 54]. We note that these tools are oriented towards penetration testing, and are designed to perform targeted scans on web properties directly under the control of the tester. That is, by design, they operate under certain pre-conditions, perform information disclosure tests via simple similarity and edit distance checks, and otherwise require manual supervision and interpretation of the results. This is orthogonal to the methodology and findings we present in this paper. Our experiment is, instead, designed to discover WCD vulnerabilities at scale in the wild, and does not rely on page similarity metrics that would result in an overwhelming number of false positives in an uncontrolled test environment.

2.4 Other Related Work

Caching mechanisms in many Internet technologies (e.g., ARP, DNS) have been targeted by cache poisoning attacks, which involve an attacker storing a malicious payload in a cache later to be served to victims. For example, James Kettle recently presented practical cache poisoning attacks against caching proxies [37, 38]. Likewise, Nguyen et al. demonstrated that negative caching (i.e., caching of 4xx or 5xx error responses) can be combined with cache poisoning to launch denial-of-service attacks [47]. Although the primary goal of a cache poisoning attack is malicious payload injection and not private data disclosure, these attacks nevertheless manipulate web caches using mechanisms similar to web cache deception. Hence, these two classes of attacks are closely related.

More generally, the complex ecosystem of CDNs and their critical position as massively-distributed networks of caching reverse proxies have been studied in various security contexts [56, 26]. For example, researchers have explored ways to use CDNs to bypass Internet censorship [29, 67, 22], exploit or weaponize CDN resources to mount denial-of-service attacks [60, 11], and exploit vectors to reveal origin server addresses behind proxies [65, 34]. On the defense front, researchers have proposed techniques to ensure the integrity of data delivered over untrusted CDNs and other proxy services [40, 42, 44]. This research is orthogonal to WCD, and is not directly relevant to our results.

3 Methodology

Figure 2: A high-level overview of our WCD measurement methodology.

We present our measurement methodology in three stages: (1) measurement setup, (2) attack surface detection, and (3) WCD detection. We illustrate this process in Figure 2. We implemented the tools that perform the described tasks using a combination of Google Chrome and Python’s Requests library [52] for web interactions, and Selenium [53] and Google Remote Debugging Protocol [25] for automation.

3.1 Stage 1: Measurement Setup

WCD attacks are only meaningful when a vulnerable site manages private end-user information and allows performing sensitive operations on this data. Consequently, sites that provide authentication mechanisms are prime targets for attacks, and thus also for our measurements. The first stage of our methodology identifies such sites and creates test accounts on them.111In the first measurement study we present in Section 4, we scoped our investigation to sites that support Google OAuth [51] for authentication due to its widespread use. This was a design choice made to automate a significant chunk of the initial account setup workload, a necessity for a large-scale experiment. In our follow-up experiment later described in Section 5 we supplemented this data set with an additional 45 sites that do not use Google OAuth. We discuss these considerations in their corresponding sections.

Domain Discovery.

This stage begins by visiting the sites in an initial measurement seed pool (e.g., the Alexa Top 

domains). We then increase site coverage by performing sub-domain discovery using open-source intelligence tools 

[50, 1, 27]. We add these newly-discovered sub-domains of the primary sites (filtered for those that respond to HTTP(s) requests) to the seed pool.

Account Creation.

Next, we create two test accounts on each site: one for a victim, and the other for an attacker. We populate each account with unique dummy values. Next, we manually explore each victim account to discover data fields that should be considered private information (e.g., name, email, address, payment account details, security questions and responses) or user-created content (e.g., comments, posts, internal messages). We populate these fields with predefined markers that can later be searched for in cached responses to detect a successful WCD attack. On the other hand, no data entry is necessary for attacker accounts.

Cookie Collection.

Once successfully logged into the sites in our seed pool, crawlers collect two sets of cookies for all victim and attacker accounts. These are saved in a cookie jar to be reused in subsequent steps of the measurement. Note that we have numerous measures to ensure our crawlers remain authenticated during our experiments. Our crawlers periodically re-authenticate, taking into account cookie expiration timestamps. In addition, the crawlers use regular expressions and blacklists to avoid common logout links on visited pages.

3.2 Stage 2: Attack Surface Detection

Domain Crawls.

In the second stage, our goal is to map from domains in the seed pool to a set of pages (i.e., complete URLs) that will later be tested for WCD vulnerabilities. To this end, we run a recursive crawler on each domain in the seed pool to record links to pages on that site.

Group By URL
Query Parameter
Path Parameter
Table 1: Sample URL grouping for attack surface discovery.

URL Grouping.

Many modern web applications customize pages based on query string or URL path parameters. These pages have similar structures and are likely to expose similar attack surfaces. Ideally, we would group them together and select only one random instance as a representative URL to test for WCD in subsequent steps.

Since performing a detailed content analysis is a costly process that could generate an unreasonable amount of load on the crawled site, our URL grouping strategy instead focuses on the structure of URLs, and approximates page similarity without downloading each page for analysis. Specifically, we convert the discovered URLs into an abstract representation by grouping those URLs by query string parameter names or by numerical path parameters. We select one random instance and filter out the rest. Table 1 illustrates this process.

This filtering of URLs significantly accelerates the measurements, and also avoids overconsumption of the target site’s resources with redundant scans in Stage 3. We stop attack surface detection crawls after collecting 500 unique pages per domain for similar reasons.

3.3 Stage 3: WCD Detection

In this final stage, we launch a WCD attack against every URL discovered in Stage 2, and analyze the response to determine whether a WCD vulnerability was successfully exploited.

WCD Attack.

The attack we mount directly follows the scenario previously described in Section 2.3 and illustrated in Figure 1. For each URL:

  1. We craft an attack URL that references a non-existent static resource. In particular, we append to the original page ‘‘/<random>.css’’222Our choice to use a style sheet in our payload is motivated by the fact that style sheets are essential components of most modern sites, and also prime choices for caching. They are also a robust choice for our tests. For instance, many CDN providers offer solutions to dynamically resize image files on the CDN edge depending on the viewport of a requesting client device. Style sheets are unlikely to be manipulated in such ways.. We use a random string as the file name in order to prevent ordinary end-users of the site from coincidentally requesting the same resource.

  2. We initiate a request to this attack URL from the victim account and record the response.

  3. We issue the same request from the attacker account, and save the response for comparison.

  4. Finally, we repeat the attack as an unauthenticated user by omitting any session identifiers saved in the attacker cookie jar. We later analyze the response to this step to ascertain whether attackers without authentication credentials (e.g., when the site does not offer open or free sign ups) can also exploit WCD vulnerabilities.

Marker Extraction.

Once the attack scenario described above is executed, we first check for private information disclosure by searching the attacker response for the markers that were entered into victim accounts in Stage 1. If victim markers are present in URLs requested by an attacker account, the attacker must have received the victim’s incorrectly cached content and, therefore, the target URL contains an exploitable WCD vulnerability. Because these markers carry relatively high entropy, it is probabilistically highly unlikely that this methodology will produce false positives.

Secret Extraction.

We scan the attacker response for the disclosure of secret tokens frequently used as part of web application security mechanisms. These checks include common secrets (e.g., CSRF tokens, session identifiers) as well as any other application-specific authentication and authorization tokens (e.g., API credentials). We also check for session-dependent resources such as dynamically-generated JavaScript, which may have private information and secrets embedded in them (e.g., as explored by Lekies et al. [39]).

In order to extract candidates for leaked secrets, we scan attacker responses for name & value pairs, where either (1) the name contains one of our keywords (e.g., csrf, xsrf, token, state, client_id), or (2) the value has a random component. We check for these name & value pairs in hidden HTML form elements, query strings extracted from HTML anchor elements, and inline JavaScript variables and constants.

Similarly, we extract random file names referenced in HTML script elements. We perform all tests for randomness by first removing dictionary words from the target string (i.e., using a list of 10,000 common English words [35]), and then computing Shannon entropy over the remaining part.

Note that unlike our checks for private information leaks, this process can result in false positives. Therefore, we perform this secret extraction process only when the victim and attacker responses are identical (a strong indicator of caching), or otherwise when we can readily confirm a WCD vulnerability by searching for the private information markers. In addition, we later manually verify all candidate secrets extracted in this step.

3.4 Verification and Limitations

Researchers have repeatedly reported that large-scale Internet measurements, especially those that use automated crawlers, are prone to being blocked or served fake content by security solutions designed to block malicious bots and content scrapers [49, 66]. In order to minimize this risk during our measurement, we used a real browser (i.e., Google Chrome) for most steps in our methodology. For other interactions, we set a valid Chrome user-agent string. We avoided generating excessive amounts of traffic and limited our crawls as described above in order to avoid triggering rate-limiting alerts, in addition to ethical motivations. After performing our measurements, we manually verified all positive findings and confirmed the discovered vulnerabilities.

Note that this paper has several important limitations, and the findings should be considered a potentially loose lower bound on the incidence of WCD vulnerabilities in the wild. For example, as described in Section 4, our seed pool is biased toward sites that support Google OAuth, which was a necessary compromise to automate our methodology and render a large-scale measurement feasible. Even under this constraint, creating accounts on some sites required entering and verifying sensitive information such as credit card or US social security numbers which led to their exclusion from our study.

Furthermore, decisions such as grouping URLs based on their structure without analyzing page content, and limiting site crawls to 500 pages may have caused us to miss additional instances of vulnerabilities. Similarly, even though we manually filtered out false positives during our secret token extraction process and verified all findings, we do not have a scalable way of detecting false negatives. We believe that these trade-offs were worthwhile given the overall security benefits of and lessons learned from our work. We emphasize that the results in this paper represent a lower bound.

3.5 Ethical Considerations

Here, we explain in detail important ethical considerations pertaining to this work and the results we present.

Performance Considerations.

We designed our methodology to minimize the performance impact on scanned sites and inconvenience imposed on their operators. We did not perform repeated or excessive automated scans of the targeted sites, and ensured that our measurements did not generate unreasonable amounts of traffic. We used only passive techniques for sub-domain enumeration and avoided abusing external resources or the target site’s DNS infrastructure.

Similarly, our stored modifications to crawled web applications only involved creating two test accounts and filling out editable fields with markers that we later used for data leakage detection. We believe this will have no material impact on site operators, especially in the presence of common threats such as malicious bots and credential stuffing tools that generate far more excessive junk traffic and data.

Security Considerations.

Our methodology entirely avoids jeopardizing the security of crawled sites or their end-users. In this work, we never injected or stored any malicious payload to target sites, to web caches on the communication path, or otherwise maliciously tampered with any technology involved in the process. Likewise, the experiments we performed all incorporated randomized strings as the non-existent parts of URLs, thereby preventing unsuspecting end-users from accidentally accessing our cached data and receiving unexpected responses.

Note that this path randomization measure was used to prevent inconveniencing or confusing end-users; since we never exploited WCD to leak real personal data from a web application or stored a malicious payload, our work never posed a security risk to end-users.

Our experiments did not take into account robots.txt files. This was a risk-based decision we consciously made, and we believe that ignoring exclusion directives had no negative impact on the privacy of these sites’ visitors. Robots.txt is not a security or privacy mechanism, but is intended to signal to data aggregators and search engines what content to index – including a directive to exclude privacy sensitive pages would actually be a misuse of this technology. This is not relevant to our experiments, as we only collect content for our analysis, and we do not index or otherwise publicly present site content.

Responsible Disclosure.

In this paper, we present a detailed breakdown of our measurement findings and results of our analysis, but we refrain from explicitly naming the impacted sites. Even though our methodology only utilized harmless techniques for WCD detection, the findings point at real-world vulnerabilities that could be severely damaging if publicly disclosed before remediation.

We sent notification emails to publicly listed security contacts of all impacted parties promptly after our discovery. In the notification letters we provided an explanation of the vulnerability with links to online resources and listed the vulnerable domain names under ownership of the contacted party. We informed them of our intention to publicly publish these results, noted that they will not be named, and advised that they remediate the issue as adversaries can easily repeat our experiment and compromise their sites. We also explicitly stated that we did not seek or accept bug bounties for these notifications.

We sent the notification letters prior to submitting this work for review, therefore giving the impacted parties reasonably early notice. As of this writing, 12 of the impacted sites have implemented mitigations.


One of the authors of this paper is affiliated with a major CDN provider at the time of writing. However, the work and results we present in this paper do not use any internal or proprietary company information, or any such information pertaining to the company’s customers. We conducted this work using only publicly available data sources and tools. Our methodology is repeatable by other researchers without access to any CDN provider internals.

4 Web Cache Deception Measurement Study

We conducted two measurement studies to characterize web cache deception (WCD) vulnerabilities on the Internet. In this first study we present in this section, the research questions we specifically aim to answer are:

What is the prevalence of WCD vulnerabilities on popular, highly-trafficked domains? (§4.2)

Do WCD vulnerabilities expose PII and, if so, what kinds? (§4.3)

Can WCD vulnerabilities be used to defeat defenses against web application attacks? (§4.3)

Can WCD vulnerabilities be exploited by unauthenticated users? (§4.3)

In the following, we describe the data we collected to carry out the study. We discuss the results of the measurement, and then consider implications for PII and important web security defenses. Finally, we summarize the conclusions we draw from the study. In Section 5, we will present a follow-up experiment focusing on advanced path confusion techniques.

4.1 Data Collection

Crawled Vulnerable
Pages 1,470,410 17,293 (1.2%)
Domains 124,596 93 (0.1%)
Sites 295 16 (5.4%)
Table 2: Summary of crawling statistics.

We developed a custom web crawler to collect the data used in this measurement. The crawler ran from April 20-27, 2018 as a Kubernetes pod that was allocated 16 Intel Xeon 2.4 GHz CPUs and 32 GiB of RAM. Following the methodology described in Section 3, we configured the crawler to identify vulnerable sites from the Alexa Top 5K at the time of the experiment. In order to scalably create test accounts, we filtered this initial measurement seed pool for sites that provide an option for user authentication via Google OAuth. This filtering procedure narrowed the set of sites considered in this measurement to 295. Table 2 shows a summary of our crawling statistics.

4.2 Measurement Overview

Alexa Ranking.

Figure 3: Distribution of the measurement data and vulnerable sites across the Alexa Top 5K.

From the 295 sites comprising the collected data set, the crawler identified 16 sites (5.4%) to contain WCD vulnerabilities. Figure 3 presents the distribution of all sites and vulnerable sites across the Alexa Top 5K. From this, we observe that the distribution of vulnerable sites is roughly proportional to the number of sites crawled; that is, our data does not suggest that the incidence of WCD vulnerabilities is correlated with site popularity.

Content Delivery Networks (CDNs).

CDN Crawled Vulnerable
Pages Domains Sites Pages Domains Sites
Cloudflare 161,140 (11.0%) 4,996 (4.0%) 143 (48.4%) 16,234 (93.9%) 72 (77.4%) 8 (50.0%)
Akamai 225,028 (15.3%) 16,473 (13.2%) 100 (33.9%) 1,059 (6.1%) 21 (22.6%) 8 (50.0%)
CloudFront 100,009 (6.8%) 10,107 (8.1%) 107 (36.3%) 2 (<0.1%) 1 (1.1%) 1 (6.2%)
Other CDNs 244,081 (16.6%) 2,456 (2.0%) 137 (46.4%) 0 (0.0%) 0 (0.0%) 0 (0.0%)
Total CDN Use 707,210 (48.1%) 33,675 (27.0%) 244 (82.7%) 17,293 (100.0%) 93 (100.0%) 16 (100.0%)
Table 3:

Pages, domains, and sites labeled by CDN using HTTP header heuristics. These heuristics simply check for unique vendor-specific strings added by CDN proxy servers.

Using a set of heuristics that searches for well-known vendor strings in HTTP headers, we labeled each domain and site with the corresponding CDN. Table 3 shows the results of this labeling. Note that many sites use multiple CDN solutions, and therefore the sum of values in the first four rows may exceed the totals we report in the last row.

The results show that, even though WCD attacks are equally applicable to any web cache technology, all instances of vulnerable pages we observed are served over a CDN. That being said, vulnerabilities are not unique to any one CDN vendor. While this may seem to suggest that CDN use is correlated with an increased risk of WCD, we point out that 82.7% of sites in our experiment are served over a CDN. A more balanced study focusing on comparing CDNs to centralized web caches is necessary to eliminate this inherent bias in our experiment and draw meaningful conclusions. Overall, these results indicate that CDN deployments are prevalent among popular sites, and the resulting widespread use of web caches may in turn lead to more opportunities for WCD attacks.

Response Codes.

Response Code Pages Domains Sites
404 Not Found 17,093 (98.8%) 82 (88.2%) 10 (62.5%)
200 Ok 205 (1.2%) 19 (20.4%) 12 (75.0%)
Table 4: Response codes observed in the vulnerable data set.

Table 4 presents the distribution of HTTP response codes observed for the vulnerable sites. This distribution is dominated by 404 Not Found which, while perhaps unintuitive, is indeed allowed behavior according to RFC 7234 [21]. On the other hand, while only 12 sites leaked resources with a 200 OK response, during our manual examination of these vulnerabilities (discussed below) we noted that more PII was leaked from this category of resource.

Cache Headers.

Header Pages Domains Sites
Expires: 1,642 (9.5%) 23 (24.7%) 13 (81.2%)
Pragma: no-cache 652 (3.8%) 11 (11.8%) 6 (37.5%)
Cache-Control: 1,698 (9.8%) 26 (28.0%) 14 (87.5%)
    max-age=, public 1,093 (6.3%) 10 (10.8%) 7 (43.8%)
    max-age= 307 (1.8%) 1 (1.1%) 1 (6.2%)
    must-revalidate, private 102 (0.6%) 1 (1.1%) 1 (6.2%)
    max-age=, no-cache, no-store 67 (0.4%) 3 (3.2%) 2 (12.5%)
    max-age=, no-cache 64 (0.4%) 4 (4.3%) 1 (6.2%)
    max-age=, must-revalidate 51 (0.3%) 1 (1.1%) 1 (6.2%)
    max-age=, must-revalidate, no-transform, private 5 (<0.1%) 3 (3.2%) 1 (6.2%)
    no-cache 5 (<0.1%) 2 (2.2%) 1 (6.2%)
    max-age=, private 3 (<0.1%) 1 (1.1%) 1 (6.2%)
    must-revalidate, no-cache, no-store, post-check=, pre-check= 1 (<0.1%) 1 (1.1%) 1 (6.2%)
All 1,698 (9.8%) 26 (28.0%) 14 (87.5%)
(none) 15,595 (90.2%) 67 (72.0%) 3 (18.8%)
Table 5: Cache headers present in HTTP responses collected from vulnerable sites.

Table 5 shows a breakdown of cache-relevant headers collected from vulnerable sites. In particular, we note that despite the presence of headers whose semantics prohibit caching—e.g., “Pragma: no-cache”, “Cache-Control: no-store”—pages carrying these headers are cached regardless, as they were found to be vulnerable to WCD. This finding suggests that site administrators indeed take advantage of the configuration controls provided by web caches that allow sites to override header-specified caching policies.

A consequence of this observation is that user-agents cannot use cache headers to determine with certainty whether a resource has in fact been cached or not. This has important implications for WCD detection tools that rely on cache headers to infer the presence of WCD vulnerabilities.

4.3 Vulnerabilities

Table 6 presents a summary of the types of vulnerabilities discovered in the collected data, labeled by manual examination.

Leakage Pages Domains Sites
PII 17,215 (99.5%) 88 (94.6%) 14 (87.5%)
    User 934 (5.4%) 17 (18.3%) 8 (50.0%)
    Name 16,281 (94.1%) 71 (76.3%) 7 (43.8%)
    Email 557 (3.2%) 10 (10.8%) 6 (37.5%)
    Phone 102 (0.6%) 1 (1.1%) 1 (6.2%)
CSRF 130 (0.8%) 10 (10.8%) 6 (37.5%)
    JS 59 (0.3%) 5 (5.4%) 4 (25.0%)
    POST 72 (0.4%) 5 (5.4%) 3 (18.8%)
    GET 8 (<0.1%) 4 (4.3%) 2 (12.5%)
Sess. ID / Auth. Code 1,461 (8.4%) 11 (11.8%) 6 (37.5%)
    JS 1,461 (8.4%) 11 (11.8%) 6 (37.5%)
Total 17,293 93 16
Table 6: Types of vulnerabilities discovered in the data.


14 of the 16 vulnerable sites leaked PII of various kinds, including names, usernames, email addresses, and phone numbers. In addition to these four main categories, a variety of other categories of PII were found to be leaked. Broad examples of other PII include financial information (e.g., account balances, shopping history) and health information (e.g., calories burned, number of steps, weight). While it is tempting to dismiss such information as trivial, we note that PII such as the above can be used as the basis for highly effective spearphishing attacks [19, 32, 30, 10].

Security Tokens.

Using the entropy-based procedure described in Section 3, we also analyzed the data for the presence of leaked security tokens. Then, we manually verified our findings by accessing the vulnerable sites using a browser and checking for the presence of the tokens suspected to have been leaked. Finally, we manually verified representative examples of each class of leaked token for exploitability using the test accounts established during the measurement.

6 of the 16 vulnerable sites leaked CSRF tokens valid for a session, which could allow an attacker to conduct CSRF attacks despite the presence of a deployed CSRF defense. 3 of these were discovered in hidden form elements used to protect POST requests, while an additional 4 were found in inline JavaScript that was mostly used to initiate HTTP requests. We also discovered 2 sites leaking CSRF tokens in URL query parameters for GET requests, which is somewhat at odds with the convention that GET requests should be idempotent.

6 of the 16 vulnerable sites leaked session identifiers or user-specific API tokens in inline JavaScript. These session identifiers could be used to impersonate victim users at the vulnerable site, while the API tokens could be used to issue API requests as a victim user.

Authenticated vs. Unauthenticated Attackers.

The methodology we described in Section 3 includes a detection step intended to discover whether a suspected WCD vulnerability was exploitable by an unauthenticated user by accessing a cached page without sending any stored session identifiers in the requests. In only a few cases did this automated check fail; that is, in virtually every case the discovered vulnerability was exploitable by an unauthenticated user. Even worse, manual examination of the failure cases revealed that in each one the crawler had produced a false negative and that in fact all of the remaining vulnerabilities were exploitable by unauthenticated users as well. This implies that WCD, as a class of vulnerability, tends not to require an attacker to authenticate to a vulnerable site in order to exploit those vulnerabilities. In other words, requiring strict account verification through credentials such as valid SSNs or credit card numbers is not a viable mitigation for WCD.

4.4 Study Summary

Summarizing the major findings of this first experiment, we found that 16 out of 295 sites drawn from the Alexa Top 5K contained web cache deception (WCD) vulnerabilities. We note that while this is not a large fraction of the sites scanned, these sites have substantial user populations as to be expected with their placement in the Alexa rankings. This, combined with the fact that WCD vulnerabilities are relatively easy to exploit, leads us to conclude that these vulnerabilities are serious and that this class of vulnerability deserves attention from both site administrators and the security community.

We found that the presence of cache headers was an unreliable indicator for whether a resource is cached, implying that existing detection tools relying on this signal may inadvertently produce false negatives when scanning sites for WCD vulnerabilities. We found vulnerable sites to leak PII that would be useful for launching spearphishing attacks, or security tokens that could be used to impersonate victim users or bypass important web security defenses. Finally, the WCD vulnerabilities discovered here did not require attackers to authenticate to vulnerable sites, meaning sites with restrictive sign-up procedures are not immune to WCD vulnerabilities.

5 Variations on Path Confusion**@@nonexistent.css@@
(a) Path Parameter**@@nonexistent.css@@
(b) Encoded Newline (\n);par1;par2**@@nonexistent.css@@
(c) Encoded Semicolon (;)**@@nonexistent.css@@
(d) Encoded Pound (#)**name=val@@nonexistent.css@@
(e) Encoded Question Mark (?)
Figure 4: Five practical path confusion techniques for crafting URLs that reference nonexistent file names. In each example, the first URL corresponds to the regular page, and the second one to the malicous URL crafted by the attacker. More generally, nonexistent.css corresponds to a nonexistent file where nonexistent is an arbitrary string and .css is a popular static file extension such as .css, .txt, .jpg, .ico, .js etc.

Web cache technologies may be configured to make their caching decisions based on complex rules such as pattern matches on file names, paths, and header contents. Launching a successful WCD attack requires an attacker to craft a malicious URL that triggers a caching rule, but also one that is interpreted as a legitimate request by the web server. Caching rules often cannot be reliably predicted from an attacker’s external perspective, rendering the process of crafting an attack URL educated guesswork.

Based on this observation, we hypothesize that exploring variations on the path confusion technique may increase the likelihood of triggering caching rules and a valid web server response, and make it possible to exploit additional WCD vulnerabilities on sites that are not impacted by the originally proposed attack. To test our hypothesis, we performed a second round of measurements fourteen months after the first experiment, in July, 2019.

Specifically, we repeated our methodology, but tested payloads crafted with different path confusion techniques in an attempt to determine how many more pages could be exploited with path confusion variations. We used an extended seed pool for this study, containing 295 sites from the original set and an additional 45 randomly selected from the Alexa Top 5K, for a total of 340. In particular, we chose these new sites among those that do not use Google OAuth in an attempt to mitigate potential bias in our previous measurement. One negative consequence of this decision was that we had to perform the account creation step entirely manually, which limited the number of sites we could include in our study in this way. Finally, we revised the URL grouping methodology by only selecting and exploiting a page among the first 500 pages when there is at least one marker in the content, making it more efficient for our purposes, and less resource-intensive on our targets. In the following, we describe this experiment and present our findings.

5.1 Path Confusion Techniques

Recall from our analysis and Table 4 that our WCD tests resulted in a 404 Not Found status code in the great majority of cases, indicating that the web server returned an error page that is less likely to include PII. In order to increase the chances of eliciting a 200 OK response while still triggering a caching rule, we propose additional path confusion techniques below based on prior work [59, 61, 62]), also illustrated in Figure 4. Note that Path Parameter in the rest of this section refers to the original path confusion technique discussed in this work.

Encoded Newline (\n).

Web servers and proxies often (but not always) stop parsing URLs at a newline character, discarding the rest of the URL string. For this path confusion variation, we use an encoded newline (%0A) in our malicious URL (see Figure 3(b)). We craft this URL to exploit web servers that drop path components following a newline (i.e., the server sees, but are fronted by caching proxies that instead do not properly decode newlines (the proxy sees As a result, a request for this URL would result in a successful response, and the cache would store the contents believing that this is static content based on the nonexistent file’s extension.

Encoded Semicolon (;).

Some web servers and web application frameworks accept lists of parameters in the URL delimited by semicolons; however, the caching proxy fronting the server may not be configured to recognize such lists. The path confusion technique we present in Figure 3(c) exploits this scenario by appending the nonexistent static file name after a semicolon. In a successful attack, the server would decode the URL and return a response for, while the proxy would fail to decode the semicolon, interpret as a resource, and attempt to cache the nonexistent style sheet.

Encoded Pound (#).

Web servers often process the pound character as an HTML fragment identifier, and therefore stop parsing the URL at its first occurrence. However, proxies and their caching rules may not be configured to decode pound signs, causing them to process the entire URL string. The path confusion technique we present in Figure 3(d) once again exploits this inconsistent interpretation of the URL between a web server and a web cache, and works in a similar manner to the encoded newline technique above. That is, in this case the web server would successfully respond for, while the proxy would attempt to cache

Encoded Question Mark (?).

This technique, illustrated in Figure 3(e), targets proxies with caching rules that are not configured to decode and ignore standard URL query strings that begin with a question mark. Consequently, the web server would generate a valid response for and the proxy would cache it, misinterpreting the same URL as example/account.php%3Fname=valnonexistent.css.

5.2 Results

Technique Pages Domains Sites
200 !200 200 !200 200 !200
Path Parameter 3,870 25,932 31 93 13 7
Encoded \n 1,653 24,280 79 76 9 7
Encoded ; 3,912 25,576 91 92 13 7
Encoded # 7,849 20,794 102 85 14 7
Encoded ? 11,282 26,092 122 86 17 8
All Encoded 11,345 31,063 128 94 20 9
Total 12,668 32,281 132 97 22 9
Table 7: Response codes observed with successful WCD attacks for each path confusion variation.

We applied our methodology to the seed pool of 340 sites, using each path confusion variation shown in Figure 4. We also performed the test with the Path Parameter technique, which was an identical test case to our original experiment. We did this in order to identify those pages that are not vulnerable to the original WCD technique, but only to its variations.

We point out that the results we present in this second experiment for the Path Parameter technique differ from our first measurement. This suggests that, in the fourteen-month gap between the two experiments, either the site operators fixed the issue after our notification, or that there were changes to the site structure or caching rules that mitigated existing vulnerabilities or exposed new vulnerable pages. In particular, we found 16 vulnerable sites in the previous experiment and 25 in this second study, while the overlap between the two is only 4.

Of the 25 vulnerable sites we discovered in this experiment, 20 were among the previous set of 295 that uses Google OAuth, and 5 among the newly picked 45 that do not. To test whether the incidence distributions of vulnerabilities among these two sets of sites show a statistically significant difference, we applied Pearson’s

test, where vulnerability incidence is treated as the categorical outcome variable and OAuth/non-OAuth site sets are comparison groups. We obtained a test statistic of 1.07 and a p-value of 0.30, showing that the outcome is independent of the comparison groups, and that incidence distributions do not differ significantly at typically chosen significance levels (i.e., p

0.05 ). That is, our seed pool selection did not bias our findings.

Technique Pages Domains Sites
Path Parameter 29,802 (68.9%) 103 (69.6%) 14 (56.0%)
Encoded \n 25,933 (59.9%) 86 (58.1%) 11 (44.0%)
Encoded ; 29,488 (68.2%) 105 (70.9%) 14 (56.0%)
Encoded # 28,643 (66.2%) 109 (73.6%) 15 (60.0%)
Encoded ? 37,374 (86.4%) 130 (87.8%) 19 (76.0%)
All Encoded 42,405 (98.0%) 144 (97.3%) 23 (92.0%)
Total 43,258 (100.0%) 148 (100.0%) 25 (100.0%)
Table 8: Vulnerable targets for each path confusion variation.
Technique Path Parameter Encoded \n Encoded ; Encoded # Encoded ?
Path Parameter - 4,390 / 26 / 7 1,010 / 5 / 4 5,691 / 11 / 3 5,673 / 12 / 3
Encoded \n 521 / 9 / 4 - 206 / 5 / 3 3,676 / 5 / 3 3,668 / 5 / 3
Encoded ; 696 / 7 / 4 3,761 / 24 / 6 - 4,881 / 9 / 2 4,863 / 8 / 0
Encoded # 4,532 / 17 / 4 6,386 / 28 / 7 4,036 / 13 / 3 - 90 / 1 / 1
Encoded ? 13,245 / 39 / 8 15,109 / 49 / 11 12,749 / 33 / 5 8,821 / 22 / 5 -
All Encoded 13,456 / 45 / 11 16,472 / 58 / 12 12,917 / 39 / 9 13,762 / 35 / 8 5,031 / 14 / 4
Table 9: Number of unique pages/domains/sites exploited by each path confusion technique. Element indicates number of many pages exploitable using the technique in row , whereas technique in column is ineffective.

Response Codes.

We present the server response codes we observed for vulnerable pages in Table 7. Notice that there is a stark contrast in the number of 200 OK responses observed with some of the new path confusion variations compared to the original. For instance, while there were 3,870 success codes for Path Parameter, Encoded # and Encoded ? resulted in 7,849 and 11,282 success responses respectively. That is, two new path confusion techniques were indeed able to elicit significantly higher numbers of successful server responses, which is correlated with a higher chance of returning private user information. The remaining two variations performed closer to the original technique.


In this experiment we identified a total of 25 vulnerable sites. Table 8 shows a breakdown of vulnerable pages, domains, and sites detected using different path confusion variations. Overall, the original path confusion technique resulted in a fairly successful attack, exploiting 68.9% of pages and 14 sites. Still, the new techniques combined were able to exploit 98.0% of pages, and 23 out of 25 vulnerable sites, showing that they significantly increase the likelihood for a successful attack.

We next analyze whether any path confusion technique was able to successfully exploit pages that were not impacted by others. We present these results in Table 9 in a matrix form, where each element shows how many pages/domains/sites were exploitable using the technique in row , whereas utilizing the technique listed in column was ineffective for the same pages/domains/sites.

The results in Table 9 confirm that each path confusion variation was able to attack a set of unique pages/domains/sites that were not vulnerable to other techniques, attesting to the fact that utilizing a variety of techniques increases the chances of successful exploitation. In fact, of the 25 vulnerable sites, 11 were only exploitable using one of the variations we presented here, but not the Path Parameter technique.

All in all, the results we present in this section confirm our hypothesis that launching WCD attacks with variations on path confusion, as opposed to only using the originally proposed Path Parameter technique, results in an increased possibility of successful exploitation. Moreover, two of the explored variations elicit significantly more 200 OK server responses in the process, increasing the likelihood of the web server returning valid private information.

We stress that the experiment we present in this section is necessarily limited in scale and scope. Still, we believe the findings sufficiently demonstrate that WCD can be easily modified to render the attack more damaging, exploiting unique characteristics of web servers and caching proxies in parsing URLs. An important implication is that defending against WCD through configuration adjustments is difficult and error prone. Attackers are likely to have the upper hand in devising new and creative path confusion techniques that site operators may not anticipate.

6 Empirical Experiments

Practical exploitation of WCD vulnerabilities depends on many factors such as the caching technology used and caching rules configured. In this section, we present two empirical experiments we performed to demonstrate the impact of different cache setups on WCD, and discuss our exploration of the default settings for popular CDN providers.

CDN Default Cached Objects Honored Headers
no-store no-cache private
Akamai Objects with a predefined list of static file extensions only.
Cloudflare Objects with a predefined list of static file extensions, AND
all objects with cache control headers public or max-age > 0.
CloudFront All objects.
Fastly All objects.
Table 10: Default caching behavior for popular CDNs, and cache control headers honored by default to prevent caching.

6.1 Cache Location

While centralized server-side web caches can be trivially exploited from any location in the world, exploiting a distributed set of CDN cache servers is more difficult. A successful WCD attack may require attackers to correctly target the same edge server that their victim connects to, where the cached sensitive information is stored. As extensively documented in existing WCD literature, attackers often achieve that by connecting to the server of interest directly using its IP address and a valid HTTP Host header corresponding to the vulnerable site.

We tested the impact of this practical constraint by performing the victim interactions of our methodology from a machine located in Boston, MA, US, and launching the attack from another server in Trento, Italy. We repeated this test for each of the 25 sites confirmed to be vulnerable in our second measurement described in Section 5.

The results showed that our attack failed for 19 sites as we predicted, requiring tweaks to target the correct cache server. Surprisingly, the remaining 6 sites were still exploitable even though headers indicated that they were served over CDNs (3 Akamai, 1 Cloudflare, 1 CloudFront, and 1 Fastly).

Upon closer inspection of the traffic, we found headers in our Fastly example indicating that a cache miss was recorded in their Italy region, followed by a retry in the Boston region that resulted in the cache hit, which led to a successful attack. We were not able to explore the remaining cases with the data servers exposed to us.

Many CDN providers are known to use a tiered cache model, where content may be available from a parent cache even when evicted from a child [20, 4]. The Fastly example above demonstrates this situation, and is also a plausible explanation for the remaining cases. Another possibility is that the vulnerable sites were using a separate centralized server-side cache fronted by their CDN provider. Unfortunately, without a clear understanding of proprietary CDN internals and visibility into site owners’ infrastructure, it is not feasible to determine the exact cache interactions.

Our experiment confirms that cache location is a practical constraint for a successful WCD attack where a distributed set of cache servers is involved, but also shows that attacks are viable in certain scenarios without necessitating additional traffic manipulation.

6.2 Cache Expiration

Web caches typically store objects for a short amount of time, and then evict them once they expire. Eviction may also take place prematurely when web caches are under heavy load. Consequently, an attacker may have a limited window of opportunity to launch a successful WCD attack until the web cache drops the cached sensitive information.

In order to measure the impact of cache expiration on WCD, we repeated the attacker interactions of our methodology with 1 hour, 6 hour, and 1 day delays. 333We only tested 19 sites out of 25, as the remaining 6 had fixed their vulnerabilities by the time we performed this experiment. We found that 16, 10, and 9 sites were exploitable in each case, respectively.

These results demonstrate that exploitation is viable in realistic attack scenarios, where there are delays between the victim’s and attacker’s interactions with web caches. That being said, caches will eventually evict sensitive data, meaning that attacks with shorter delays are more likely to be successful. We also note that we performed this test with a randomly chosen vulnerable page for each site as that was sufficient for our purposes. In practice, different resources on a given site may have varying cache expiration times, imposing additional constraints on what attacks are possible.

6.3 CDN Configurations

Although any web cache technology can be affected by WCD, we established in Section 4.2 that CDNs play a large role in cache use on the Internet. Therefore, we conducted an exploratory experiment to understand the customization features CDN vendors offer and, in particular, to observe their default caching behavior. To that end, we created free or trial accounts with four major CDN providers: Akamai, Cloudflare, CloudFront, and Fastly. We only tested the basic content delivery solutions offered by each vendor and did not enable add-on features such as web application firewalls.

We stress that major CDN providers offer rich configuration options, including mechanisms for site owners to programmatically interact with their traffic. A systematic and exhaustive analysis of CDN features and corresponding WCD vectors is an extremely ambitious task beyond the scope of this paper. The results we present in this section are only intended to give high-level insights into how much effort must be invested in setting up a secure and safe CDN environment, and how the defaults behave.


All four CDN providers we experimented with offer a graphical interface and APIs for users to set up their origin servers, apply caching rules, and configure how HTTP headers are processed. In particular, all vendors provide ways to honor or ignore Cache-Control headers, and users can choose whether to strip headers or forward them downstream to clients. Users can apply caching decisions and time-to-live values for cached objects based on expressions that match the requested URLs.

Akamai and Fastly configurations are translated to and backed by domain-specific configuration languages, while Cloudflare and CloudFront do not expose their back-end to users. Fastly internally uses Varnish caches, and gives users full control over the Varnish Configuration Language (VCL) that governs their setup. In contrast, Akamai appears to support more powerful HTTP processing features than Varnish, but does not expose all features to users directly. Quoting an Akamai blog post: “Metadata [Akamai’s configuration language] can do almost anything, good and bad, which is why WRITE access to metadata is restricted, and only Akamai employees can add metadata to a property configuration directly.” [2]

In addition to static configurations, both Akamai and Cloudflare offer mechanisms for users to write programs that execute on the edge server, and dynamically manipulate traffic and caches [14, 3].

In general, while Cloudflare, CloudFront, and Fastly offer free accounts suitable for personal use, they also have paid tiers that lift restrictions (e.g., Cloudflare only supports 3 cache rules in the free tier) and provide professional services support for advanced customization. Akamai strictly operates in the business-to-business market where configuration is driven by a professional services team, as described above.


Next, we tested the caching behavior of CDN providers with a default configuration. Our observations here are limited to 200 OK responses pertaining to WCD; for an in-depth exploration of caching decisions involving 4xx or 5xx error responses, we refer readers to Nguyen et al. [47]. We summarize our observations in Table 10, which lists the conditions for caching objects in HTTP responses, and whether including the relevant Cache-Control headers prevent caching.

These results show that both Akamai and Cloudflare rely on a predefined list of static file extensions (e.g., .jpg, .css, .pdf, .exe) when making cacheability decisions. While Cloudflare allows origin servers to override the decision in both directions via Cache-Control headers, either to cache non-static files or prevent caching static files, Akamai’s default rule applies unconditionally.

CloudFront and Fastly adopt a more aggressive caching strategy: in the absence of Cache-Control headers all objects are cached with a default time-to-live value. Servers behind CloudFront can prevent caching via Cache-Control headers as expected. However, Fastly only honors the private header value.

6.4 Lessons Learned

The empirical evidence we presented in this section suggests that configuring web caches correctly is not a trivial task. Moreover, the complexity of detecting and fixing a WCD vulnerability is disproportionately high compared to launching an attack.

As we have seen above, many major CDN vendors do not make RFC-compliant caching decisions in their default configurations [21]. Even the more restrictive default caching rules based on file extensions are prone to security problems; for example, both Akamai and Cloudflare could cache dynamically generated PDF files containing tax statements if configured incorrectly. On the other hand, we do not believe that these observations implicate CDN vendors in any way, but instead emphasize that CDNs are not intended to be plug & play solutions for business applications handling sensitive data. All CDNs provide fine-grained mechanisms for caching and traffic manipulation, and site owners must carefully configure and test these services to meet their needs.

We reiterate that, while CDNs may be a prominent component of the Internet infrastructure, WCD attacks impact all web cache technologies. The complexity of configuring CDNs correctly, the possibility of multi-CDN arrangements, and other centralized caches that may be involved all imply that defending against WCD requires site owners to adopt a holistic view of their environment. Traditional security practices such as asset, configuration, and vulnerability management must be adapted to take into consideration the entire communication infrastructure as a system.

From an external security researcher’s perspective the challenge is even greater. As we have also discussed in the cache location and expiration experiments, reasoning about a web cache system’s internals in a black box fashion is a challenging task, which in turn makes it difficult to pinpoint issues before they can be exploited. In contrast, attackers are largely immune to this complexity; they often do not need to disentangle the cache structure for a successful attack. Developing techniques and tools for reliable detection of WCD—and similar web cache attacks—is an open research problem. We believe a combination of systems security and safety approaches would be a promising research direction, which we discuss next as we conclude this paper.

7 Discussion & Conclusion

In this paper, we presented the first large-scale investigation of WCD vulnerabilities in the wild, and showed that many sites among the Alexa Top 5K are impacted. We demonstrated that the vulnerable sites not only leak user PII but also secrets that, once stolen by an attacker, can be used to bypass existing authentication and authorization mechanisms to enable even more damaging web application attack scenarios.

Alarmingly, despite the severity of the potential damage, these vulnerabilities still persist more than two years after the public introduction of the attack in February 2017. Similarly, our second experiment showed that in the fourteen months between our two measurements, only 12 out of 16 sites were able to mitigate their WCD vulnerabilities, while the total number of vulnerabilities rose to 25.

One reason for this slow adoption of necessary mitigations could be a lack of user awareness. However, the attention WCD garnered from security news outlets, research communities, official web cache vendor press releases, and even mainstream media also suggests that there may be other contributing factors. In fact, it is interesting to note that there exists no technology or tool proposed to date that allows site operators to reliably determine if any part of their online architecture is vulnerable to WCD, or to close their security gaps. Similarly, there does not exist a mechanism for end-users and web browsers to detect a WCD attack and protect themselves. Instead, countermeasures are largely limited to general guidance by web cache vendors and CDN providers for their users to configure their services in consideration of WCD vectors, and the tools available offer limited manual penetration-testing capabilities for site operators with domain-specific knowledge.

We assert that the above is a direct and natural consequence of the fact that WCD vulnerabilities are a system safety problem. In an environment with WCD vulnerabilities, there are no isolated faulty components; that is, web servers, load balancers, proxies, and caches all individually perform the functionality they are designed for. Similarly, determining whether there is human error involved and, if so, identifying where that lies are both non-trivial tasks. In fact, site operators often have legitimate needs to configure their systems in seemingly hazardous ways. For example, a global corporation operating hundreds to thousands of machines may find it technically or commercially infeasible to revise the Cache-Control header settings of their individual web servers, and may be forced to instruct their CDN provider to perform caching based purely on file names.

These are all strong indicators that the growing ecosystem of web caches, in particular CDN-fronted web applications, and more generally highly-distributed Internet-based architectures, should be analyzed in a manner that captures their security and safety properties as a system. As aforementioned, venerable yet still widely-used root cause analysis techniques are likely to fall short in these efforts, because there is no individual system component to blame for the failure. Instead, security researchers should adopt a systems-centric security analysis, examining not only individual system components but also their interactions, expected outcomes, hazardous states, and accidents that may result. Modeling and analyzing WCD attacks in this way, drawing from the rich safety engineering literature [41] is a promising future research direction that will help the security community understand and address similar systems-level attacks effectively.


We thank our shepherd Ben Stock and the anonymous reviewers; this paper is all the better for their helpful feedback. This work was supported by the National Science Foundation under grant CNS-1703454, Secure Business Austria, ONR project “In-Situ Malware Containment and Deception through Dynamic In-Process Virtualization,” and EU H2020-SU-ICT-03-2018 Project No. 830929 CyberSec4Europe.