As targeted attacks gain more attention, and protection developers pay more attention to the implementation of new defensive technologies, the need arises for the testing of product efficacy with respect to this new kind of threat. However, compared with general product testing, APTs present additional challenges for the testers. In their VB2015 paper, Simon Edwards, Richard Ford and Gabor Szappanos ask if APT protection can be tested, and if so, whether it can be done practically.
Copyright © 2016 Virus Bulletin
Anyone watching the cybersecurity marketplace will have noticed a rapid rise in products that claim to provide protection against ‘Advanced Persistent Threats’ (APTs). As targeted attacks gain more attention, and protection developers pay more attention to the implementation of new defensive technologies, the need arises for the testing of product efficacy with respect to this new kind of threat. However, compared with general product testing, APTs present additional challenges for the testers. In this presentation, we ask if APT protection can be tested, and if so, whether it can be done practically.
In this paper, we address some of the challenges related to the testing of APT protection software suites and devices. Our arguments are essentially threefold. We first look at the subjective and confused range of definitions of what an APT even comprises.
We then look at some of the objections raised to testers’ measurements of APT protection efficacy in light of these definitions. Finally, we offer some simple guidelines for those who are attempting to construct or interpret tests of APT protection.
Our conclusion is that, while the entire APT space suffers at the hands of definitional uncertainty, there is a workable way forward for tests that measure different aspects of APT protection. With sufficient effort, tests can measure end-to-end protection. Such tests are ambitious but, given the importance and cost of both potential APT breaches and APT defences, a more scientific approach must be taken.
Although the term ‘APT’ is used commonly nowadays, there is no generally accepted definition for it, and this contributes greatly to the problem of testing.
In part, the APT has become this year’s buzzword, but vendors, reviewers and users employ the term differently depending on circumstance and goal. Such definitional challenges only add to the confusion.
For example, TechTarget uses the following definition:
‘An advanced persistent threat (APT) is a network attack in which an unauthorized person gains access to a network and stays there undetected for a long period of time.’ 
According to this definition the sample must be undetected to be an APT. If a product detects a threat, then it is not an APT.
This leads us to the inevitable outcome that the only valid outcome of a test of APT protection is that nothing is detected (otherwise the test sample is not an APT).
While this definition therefore significantly simplifies APT testing in general, it would make APT testing a very simple (non-existent) task so we should aim for a more practical one. Here are a few more definitions that are quite interesting:
‘APT is a set of stealthy and continuous computer hacking processes, often orchestrated by human(s) targeting a specific entity. APT usually targets organizations and/or nations for business or political motives. APT processes require a high degree of covertness over a long period of time.’ 
NSS Labs adopts an alternative acronym for a targeted attack, referring to a Targeted Persistent Attack (TPA).
‘Targeted: The attacker selected the organization, for a specific reason.
Persistent: The attack is capable of using multiple command-and-control channels and attack vectors, and constantly increasing its penetration of your IT systems and resources. It is also stubborn, resisting remediation attempts.
Attack: While the word ‘threat’ is somewhat nebulous when used in the context of APT, there is nothing unclear about it here. This is a true attack, and it may have several distinct stages.’ 
‘Advanced threat – any attack that gets past your existing defences.
Persistent threat – any successful attack that goes undetected and continues to cause damage.
Advanced persistent threat – any attack that gets past your existing defences, goes undetected and continues to cause damage.’ 
The problem with these definitions is, once again, that they attribute being undetected to being a core feature of an APT. This definition renders APT defences and tests useless. Other definitions focus on other aspects:
‘An Advanced Persistent Threat (APT) is a targeted attack against a high-value asset or a physical system.’ 
While this is a useful definition that makes it easy to determine if an attack belongs to this category, it does not explain the significance of the ‘Advanced’ and ‘Persistent’ attributes of an APT.
‘Advanced Persistent Threats (APTs) are a cybercrime category directed at business and political targets. APTs require a high degree of stealithiness [sic] over a prolonged duration of operation in order to be successful…
Advanced – Criminal operators behind the threat utilize the full spectrum of computer intrusion technologies and techniques…
Persistent – Criminal operators give priority to a specific task, rather than opportunistically seeking immediate financial gain… Threat – means that there is a level of coordinated human involvement in the attack, rather than a mindless and automated piece of code…’ 
We have a lot of definitions that attempt to define APT on an abstract level, hardly helping testers to categorize test scenarios. Our best option at this point is to change scope and deal with a better defined and more practical definition of targeted attacks along the lines of the RSA definition.
The terms ‘APT’ and ‘targeted attack’ are often used synonymously by the press and the APT protection providers so it makes sense to stick to the easily definable ‘targeted attack’ cases in test scenarios.
For practical purposes of testing we will define targeted attacks as follows:
A targeted attack is an infection scenario executed against a limited and pre-selected set of high-value assets or physical systems with the explicit purpose of data exfiltration or damage.
Given that we know now what an APT is for all practical purposes, we have still to explore what such a targeted attack might look like.
According to Mandiant , an APT attack is more of a campaign than a single event, which follows the following rough outline:
Reconnaissance – prior to performing the attack, information is gathered that is used as part of the social engineering repertoire during the later stages.
Initial compromise – performed by use of social engineering and spear phishing, over email and/or by planting malware on a website that the victim employees are likely to visit.
Establish foothold – plant remote administration software in the victim’s network to create network backdoors and tunnels allowing stealth access to its infrastructure.
Escalate privileges – use exploits and password cracking to acquire administrator privileges over the victim’s computer and possibly expand it to Windows domain administrator accounts.
Internal reconnaissance – collect information on the surrounding infrastructure, trust relationships and Windows domain structure.
Move laterally – expand control to other workstations, servers and infrastructure elements and perform data harvesting on them.
Maintain presence – ensure continued control over access channels and credentials acquired in previous steps.
Complete mission – exfiltration of stolen data from the victim’s network.
For the illustration of a targeted attack workflow, we picked up an attack that exploited the popular word processing software Microsoft Word.
This attack scenario happens in many consecutive steps but does not involve some of the steps in the previous list (initial reconnaissance, lateral movement). Regardless, it gives a good technical overview of what happens in the background on an attacked system.
Phishing email: The malware is delivered in email.
Exploited document: The email attachment is a .DOC document that exploits multiple Microsoft Office vulnerabilities. It was generated by the Microsoft Word Intruder exploit generator toolkit.
Upon opening the document, one of the vulnerabilities is triggered (depending on the patch state of the system) and the ‘shellcode’ that is embedded in the document is executed.
Shellcode: Although each of the vulnerabilities (CVE-2012-0158, CVE-2013-3906 and CVE-2014-1761) has an exploit block and a separate shellcode, they all do the same thing (see Figure 1).
The shellcode searches for the embedded payload code in the memory, brute-forcing through all readable memory pages. If this embedded code is found, it decrypts and executes the payload, which is encrypted with a simple one-byte XOR algorithm.
To make the encryption key less obvious, the 0 bytes and the key bytes are left intact. If it was not done this way then the large blocks of zero bytes that are normally present in PE files would convert to the value of the key byte and simply looking in the file for a large block of similar bytes would reveal the encryption key.
Downloader code: The decrypted payload code is not a Windows trojan. Instead, it is another piece of shellcode that downloads the final payload component. The shellcode then connects back to the C&C server and downloads and executes the required component from there (see Figure 2).
Connect-back code: In parallel to the exploitation and the downloading process the malicious document connects back to a C&C server to report the infection attempt. This way the attackers can keep track of the targeted computer, even if the infection is not ultimately successful (see Figure 3).
Final payload: The final payload is a keylogger trojan that collects keystrokes and harvests user account credentials from the gathered data. Clipboard content is also copied. The stolen data is uploaded to the attacker’s FTP server (see Figure 4).
Layered defences against targeted attacks: The attack described above has multiple steps. A modern anti-malware solution typically features several modules that utilize different detection approaches to provide a multi-layered diverse solution. This makes it more difficult for malware to bypass defences.
These modules could include the following:
Application control: blocks the execution of potentially unwanted/unauthorized applications
Anti-spam: blocks bulk email
Scanner: specific detection for known malware; generic detection for new malware
Firewall: blocks outbound communication attempts and inbound attacks
IPS: Intrusion Prevention System packet-level filtering of network traffic
URL filtering: reputation or blacklist-based blocking of website addresses
DLP: Data Loss Prevention software prevents exfiltration of sensitive data
Exploit protection: detects exploitation of application vulnerabilities
Behavioural-based detection: detects malware based on runtime activities in the system.
These different methods provide defence in depth against targeted attack scenarios, giving opportunities to detect and block the scenario at multiple points. In the example detailed earlier, the different modules can block the scenario at several points, as illustrated in Figure 5.
Any test that uses only a subset of the aforementioned detection capabilities is incomplete and, as such, does not give a full picture of the protection capabilities of the assessed solution.
As an example, in a test in which only the VirusTotal detection result of the final payload binary is used, several of the multi-layered defence modules are ignored. This is illustrated with dimmed layers in Figure 6.
Thus, testing end-to-end protection is critical if a broader view of product efficacy is to be gained.
There are a number of reasons why vendors of anti-APT products may object to testers evaluating them. There are public statements criticizing tests in the press, and earlier in this paper we explored some definitions of APTs that clearly show vendors, analysts and others having very different opinions on what an APT actually is.
If testers and vendors can’t agree on what products are supposed to do, then this brings a major obstacle to testing. Here are some reasons why vendors may object to testers evaluating their products:
Historic skepticism about security testing/testers in general.
The APT detection/protection market is young and very sensitive to test results.
Tests are too limited to provide useful results because they are not based on what is happening in the ‘real world’.
Claims that products are not testable because a realistic test requires:
defenders to react to alerts.
malware/exploits that will always bypass other defensive technologies.
‘Testers know nothing about advanced threats, how enterprises manage their networks and are too focused on traditional anti-virus testing methods to be able to test anti-APT products properly.’
This theoretical objection to anti-APT product testing may be the result of a reflexive response to standard methods of testing anti-virus software and the reputation of well-known anti-virus testers.
Some who today work in anti-APT businesses used to work in the anti-virus industry, which is a notoriously small world. It is possible that these people may be prone to assuming (based on historic experience) that anti-virus products do nothing more than scanning files to generate signatures, which are then compared to white/blacklists.
This definition of anti-virus was probably in the mind of the Symantec employee when he told the Wall Street Journal in May 2014 that ‘anti-virus is dead’.
From a business perspective it also suits anti-APT businesses to define anti-virus as being mainly or wholly signature-based because they can then demonstrate that their more sophisticated product(s) provide better protection.
Former anti-virus people probably remember the dissatisfaction there is/was with anti-virus tests that simply scan files rather than executing them or, even better, downloading them from live sources in real time. It took anti-malware testers a long time to update their methodologies, which damaged testers’ reputations in general and not just the reputations of the testers who refused to change. This damage may persist in the some minds.
If vendors assume that anti-virus means scanning files and comparing signatures, and that testers limit themselves to using on-demand scanning solely as a methodology, then it’s not surprising that they would object to such traditional and limited assessments of their sophisticated products. If vendors believe that testers are dishonest in representing the results then their objections seem even more reasonable.
However, today anti-virus products include a much wider set of protection features and many testers have adopted relevant testing methodologies to take these into account. Some also take very open approaches to disclosing methodologies and test data.
Of course, it may not suit anti-APT vendors to acknowledge this progress in the testing world because if they do, then tests gain credibility. Without credible test results vendors can make uncontested claims about products.
The advanced threat detection industry is relatively new, in terms of how it markets itself. In the following example we examine FireEye’s positioning and technology, but the same argument could be made for any number of new vendors in this space.
According to marketing material, FireEye offers what sounds like a novel proposition to detect or prevent advanced attacks. The vendor has apparently been successful in marketing its approach.
Whether or not FireEye’s combination of sandboxes and signature-based anti-virus scanners is very different to offerings from other security vendors doesn’t much matter currently because it is perceived by the market to be different.
FireEye makes a clear effort to differentiate itself from more ‘traditional’ companies and includes quotes on its website such as the following (with our emphasis):
‘FireEye is clearly the next generation of network security. Their unique signature-less technology is simply unmatched in detecting and blocking this new breed of advanced malware. (Director of security - global financial services company)’
The emphasized key phrases clearly show an effort to distinguish FireEye’s approach as being different from traditional anti-virus technology. The statements claim a lack of reliance on file signatures (even though the product actually includes an anti-virus scanner) and imply that there is a new type of malware that will always evade other solutions, in total contrast to FireEye’s solution.
At this early stage in the game positive test results would help these new companies build market share and stock price, while negative results are almost certain to damage the companies fast and hard. These ‘next-generation’ products are largely unproven and so are vulnerable to doubt from potential customers and investors. As such it may be better, from the vendors’ point of view, for there to be no tests running at all. As proposed above, a lack of results allows for uncontested marketing claims.
If tests exist that provide a vendor with results that they find acceptable for marketing purposes, then they may resist any evolution of testing procedures that could challenge their successful image. The winning vendor(s) may not care that those tests lack sophistication because their marketing objective (good results) is achieved.
Anti-APT products are complex and involve monitoring real network traffic, servers and other endpoints for attacks. As such one might expect a realistic test to completely reproduce similar networks and attacks. This is unlikely to be feasible for even the most well-resourced testing organization.
That being the case, FireEye has concluded that testing is impossible and that the best solution is simply to buy its solution. FireEye CTO Dave Merkel told CRN that, ‘the best way to evaluate FireEye is for organizations to deploy our technology in their own environment and they will understand why we are the market leader in stopping advanced attacks.’
But what is the ‘real world’ when it comes to threats and protection against those threats? Is it possible to at least test some very important parts of anti-APT products? Or should testers give up and leave customers simply to trust vendor claims?
To continue with FireEye’s statement, Merkel mentioned ‘sophisticated criminal networks and nation states’ as being the perpetrators of APT attacks. This leads us to two important questions:
Is an APT attack the exclusive domain of ultra-sophisticated and extremely well-resourced actors?
Is the attack only an ‘APT’ if those actors always use the most technically advanced attack techniques? Does it lose its ‘APT’ label if the same actors use less sophisticated tools, tactics and techniques?
If the answer to both of the above questions is ‘yes’ then a government intelligence agency could target competitors using relatively unsophisticated techniques and, in doing so, would evade the ‘APT’ classification and, presumably, the anti-APT products on offer.
A ‘yes’ answer would also lead us to conclude that opponents of an organization that are neither part of a criminal organization nor a government-sponsored group could use extremely advanced tools and tactics to steal information but would be able to evade the ‘APT’ classification and, again presumably, the anti-APT products/services.
Of course, if APTs are the exclusive domain of spies and the Mafia then, regardless of how sophisticated they are, testers will never be able to launch an APT and the testing game is over. That is unless a tester becomes a spy or master criminal… Let’s have a sanity check:
A national intelligence agency sets up a watering hole attack using a commercial exploit toolkit in order to gain long-term access to political activists. It uses the toolkit to appear less sophisticated for disavowal purposes when the attack is inevitably uncovered in time.
An individual discovers a software vulnerability and develops a zero-day exploit, which s/he uses to gain and maintain access to a network. This access is used to create deeper levels of access to systems on the network, which is then used to steal data on a regular basis. The data may be of interest to that individual for a number of reasons, to sell or use directly.
Are neither of the above examples APTs? If not, why not? Customers of anti-APT products would most likely not care about subtle labelling. They would expect both of the above events to be detected and ultimately curtailed, if not prevented in the first place.
When a vendor becomes aware of a report containing results related to its own product it frequently needs to spend time and other resources on analysing the report and associated data, and then possibly responding to any criticism it may believe to be unfair.
Testers may even request that the vendor provides the necessary equipment and expertise in setting up the product for testing. Vendors who ignore such requests risk the tester obtaining the incorrect product, setting it up incorrectly and misunderstanding how it is supposed to work.
In some cases, notably with vendors that provide managed services, it is necessary to involve staff to provision the service. This may even involve a partner company. The expense of provisioning, monitoring and reacting to the test can be significant.
This is a genuinely tricky element of any enterprise-level technology test. In a real working situation, anti-APT products do not mitigate all threats automatically. There most likely has to be human intervention from the defender, such as an administrator noticing and acting on alerts.
How this person or persons behave will vary depending on their resources, such as the level of their training, their technical skills, commitment, and the amount of time they have available to analyse and react to alerts.
Testers need to replicate the behaviour of one or more operators of an anti-APT solution. For useful results it may require multiple tests with different operator behaviours. This could turn into a behavioural test more akin to a social studies project than the technical test most would expect.
If the right type of malware and exploits for an anti-APT test have to be unknown then how is it possible for a tester to conduct such a test? Possible methods might include creating new threats or discovering threats that are live on the Internet that are not detected by any known anti-malware product.
Creating new malware requires a set of skills not usually found in anti-malware testing labs, although such tests have been known. For example, Simon P.G .Edwards has performed consumer computer magazine tests in the past that used heavily modified trojans, while MRG Effitas regularly conducts tests using ‘simulators’ (modified trojans) and even live botnets.
However, in all of these cases the threats are at least based on known malware and it is unlikely that there are many testers competent enough to create complete new and sufficiently sophisticated malware. It may not be necessary to do so, though.
Tests should be able to assess claims made about the anti-APT products, which are supposed to be able to handle advanced attacks and malware. How advanced these elements should be will depend on how advanced they are in real life. Recent reports suggest that APTs in the real world often compromise a chain of events and threats that are quite easily copied by testers.
For example, in September 2014 FireEye published a blog post detailing an APT campaign by the so-called APT12 ‘cyber espionage’ group . The tools, tactics and techniques used should be familiar to most in the anti-malware testing industry and are not especially sophisticated in terms of unknown malware and exploits.
In fact, the campaign in question used an exploit toolkit to attack a two year-old vulnerability (CVE-2012-0158). The tactics used involved attaching an ‘infected’ Word document to a spear-phishing email. This is very similar to the more advanced attack described in Section 3.1.
It does not seem that this particular campaign was so advanced that something similar could not be run in a test.
This objection, which stems from FireEye’s main marketing message at the time of writing, is quite clever because it is sort of meaningless but also essentially excludes any malware that can be detected by any other anti-malware solution. This in turn raises the bar for testers to the point where they have to find completely new malware that has never been encountered by any known security vendor.
The criterion is also quite meaningless, though, because it’s not so much the case that malware can bypass protection mechanisms but more that protection mechanisms may fail to detect or protect against the malware. This sounds like semantics so let’s explore what this subtlety really means.
Malware is not usually capable of ‘bypassing’ an anti-virus product. A dropped file would not, before it executes, usually be capable of very much at all. Anti-virus products can, however, miss malware.
Let’s imagine a malware binary that is not known to any anti-malware vendor. This would fit the above criteria for inclusion in a test if it could be proven to be universally undetectable. (Let us, for the minute, forget how hard it would be to confirm for sure that no anti-malware product in the world could detect it without alerting anti-malware vendors as to the sample’s existence.)
In another example let’s consider a malicious program that is detectable by many popular anti-malware products but not by Microsoft System Center Endpoint Protection. Does this sample fit the testing criteria? After all, it can bypass at least one solution so maybe it is valid to use this as part of a test attack. This is where things become meaningless. Does the malware need to be undetectable by one, four or all other solutions before it becomes a valid candidate for a test? How unknown does it need to be, to be unknown?
In the second example, maybe it is unknown only to Microsoft today. Does it become ineligible for inclusion in the test when sample-sharing systems catch up and the data reaches Microsoft?
A vendor might complain that testers should not assess its product using regular malware because that’s not what it’s supposed to protect against. But it’s fair to test a car’s safety features by driving it into a wall because, although you’re not supposed to, people want to know what happens when such events happen. It doesn’t mean that the product is useless if it misses some regular malware that should be picked up by another solution, and it doesn’t mean that the car is useless if it crumples under the impact of the collision.
If customers expect a SourceFire, FireEye, Fortinet or Palo Alto product to protect the network all the time by blocking new threats they may be disappointed to learn that these products do not offer this as their main service, although this is not to say they cannot still effectively defend against an APT.
Most such products monitor the network and, after detecting a threat that initially evaded them, can correlate data to discover what happened in the past.
The term ‘block’ is used a lot in marketing material and implies initial blocking, whereas in fact it may refer to blocking the threat after it has existed on the network for some time. This is an important factor when formulating a suitable testing methodology.
Table 1 includes a list of some of the obstacles that we (informally) hear for APT defence testing. We offer two options for overcoming these roadblocks. The first takes a practical approach that addresses the roadblock while the second is a more encompassing approach that illustrates the ‘high-end’ solution that is probably financially impractical but which would represent a ‘gold standard’ in APT testing.
|Obstacle||Solution #1 (practical)||Solution #2 (ambitious)|
|All tests will not be real-world.||Use same tools, tactics and techniques as adversaries use in real-world APT campaigns.||Testers administer multiple defence systems on real target networks.|
|Attackers will eventually gain access if persistent so testing becomes a penetration test that necessarily requires the reaction of the target’s IT staff.||Assess capabilities of automated elements and assume/categorize levels of skill, commitment and resources of the target. There may be more than one set of results for each product/attack. Assess usability and thoroughness of alert and logging system.||Perform a ‘Capture the Flag’ contest with independent penetration testers as the attackers and vendor-supplied consultants as the defenders (who are under pressure to perform some duties unrelated to monitoring logs, perhaps).|
|Testers require a range of attack skills beyond sourcing and executing malware from various sources.||Train on a range of attack skills or create bespoke environments with known weaknesses.||Sub-contract attacks to professional penetration testers.|
Table 1. APT defence testing obstacles.
Based on our discussion it is clear that there is a continuum of possible attacks, all of which fit someone’s definition of what an APT is. Thus tests will, by necessity, reflect this spectrum where, as we put it, the attacker’s skills range ‘from zero to Neo’.
Be clear on the APT actually being tested.
Based on this discussion, our first requirement for any APT test is that the definition of APT is clearly stated and the threat model is described.
Beyond this, it is important that the tester explains what is being tested. Is the tester claiming that the test covers the entire lifecycle of the APT or does it handle just one or two steps? If the former then scoring becomes a huge issue. Prima facie, the earlier in the lifecycle that the threat is detected the better. For tests that attempt to cover the range of detection options offered by anti-APT suites examine carefully how the scores are applied at different layers.
Be clear on whether the threat is zero, Neo or other.
We still see attacks using two-year-old exploits and we have threats such as Stuxnet that are loaded up with zero-day attacks. When looking at a test of APT solutions make sure that the tester is clear on what level of protection is offered and from what attack. Is the test against custom malware or is it simply a Metasploit-generated year-old exploit?
While the idea of testing with zero-day attacks sounds impossible, in fact there is a way in which such a test can be done. Instead of finding zero-day vulnerabilities in existing code, insert new vulnerabilities into open-source software and then build exploits for these weaknesses. This approach has the advantages that it creates no new threat to the general public (the exploit only works for our custom applications) and it tests a product’s ability to mitigate unknown threats.
Be clear on whether this is a test of a layer or a suite.
One of the mistakes it is easy to make when evaluating APT protection is to reduce the protection range and to focus on just a single aspect of protection, such as exploit detection. While such tests are valid it is important that we do not reduce the problem to just one layer of protection. For example, consider hypothetical products A and B.
Product A has outstanding exploit detection, but little else. Product B has a wide range of protection techniques, none of which is perfect but, when combined, these layers provide an extremely solid prevention and detection platform. Given that APTs can come in many forms (and some do not even have to use exploits) B probably has better protection than A in general. However, for some types of customer Product B reflects the ‘best’ component in their home-built layered protection scheme.
The guidance here is simple: be sure you understand what rolled up scores represent and do not blindly accept rankings that may not measure what you want. Furthermore, tests that claim to be assessing APT protection must consider all layers or fail to be a valid test of what it claims (although it may be a valid test of specific functionality).
Any APT test must examine infiltration.
We argue that the earlier in the attack chain that the APT is caught the better. Thus, for a layered test, it is important that testers examine the most common infiltration mechanisms. This would include, at a minimum, email, web, exploit, offline (media), social engineering, file execution and downloader/droppers, which is quite a list.
APT testing is complicated enough without having to consider the fact that the term APT is ill-defined. Furthermore, in a real APT scenario the attacker will not blindly throw in the towel when an attack is blunted, but will probe until a weakness is found. This co-evolutionary aspect between attacker and defender is most obvious when considering targeted attacks, as the attackers have specific goals and desires and will adapt to defences as best as they are able. Stopping the first salvo in an APT attack is not the end of the road. Instead, the protective countermeasures must withstand multiple waves of attack.
We argue that the single most important step with respect to moving forward is for testers and test consumers to be clear about the purpose of the test and whether that test can measure the feature(s) in question. This one single question can help both in test creation and interpretation.
The authors note that the views expressed in this paper are their own personal opinions and do not necessarily reflect the views of either AMTSO or their current employers.
 Advanced Persistent Threat (APT). TechTarget. November 2010. http://searchsecurity.techtarget.com/definition/advanced-persistent-threat-APT.
 Advanced Persistent Threat. Wikipedia. http://en.wikipedia.org/wiki/Advanced_persistent_threat.
 The Targeted Persistent Attack (TPA). NSS Labs. 19 August 2012. https://www.nsslabs.com/blog/targeted-persistent-attack-tpa-when-thing-goes-bump-night-really-bogeyman.
 Defining the “Advanced Persistent Threat”. Gartner. 11 November 2010. http://blogs.gartner.com/john_pescatore/2010/11/11/defining-the-advanced-persistent-threat.
 Juels, A.; Yen, T.-F. Sherlock Holmes and the Case of the Advanced Persistent Threat. RSA, 2012. https://www.usenix.org/conference/leet12/workshop-program/presentation/juels.
 Advanced Persistent Threats: A Brief Description. Damballa. https://www.damballa.com/advanced-persistent-threats-a-brief-description.
 Phases of an APT attack. Mandiant. http://intelreport.mandiant.com.
 Darwin’s Favorite APT Group. FireEye. https://www.fireeye.com/blog/threat-research/2014/09/darwins-favorite-apt-group-2.html.