Since the infamous hack of RSA Security’s EMC division in 2011 the traditional IT security industry has been quite clear – protecting ourselves comprehensively against persistent threats and malware has become almost impossible. Until now that is.
Ziften CEO Mike Hamilton explains the new developments in the use of Artificial Intelligence and particularly the application of machine learning to endpoint security.
Securing the perimeter is the age-old problem in the enterprise IT world, as the rise of networks, the growth of the Internet of Things (IoT) and its billions of connected devices and software applications running on servers is leading to an even wider and deeper array of vulnerabilities.
Even one successful attack can now mean that once hacked, it can cost companies, educational institutions and government agencies millions of dollars in lost time and burned up resources just to remediate the damage done.
New malware in the 1980s was appearing at the alarming rate of 30K pieces per day. The market began to respond with a large armada of application firewalls and anti-virus (AV) protection in the early 1990’s—using lists of known threat signatures to stop attacks in their tracks. However, these rudimentary solutions were far less dynamic than the cyber hackers themselves, and the thieves were winning.
The Cyber Kill Chain was created by Lockheed Martin as a guide to the basic framework of a threat that included all the necessary stages of an attack would have to go through to be successful.
The model represents a simple outline of the anatomy of a cyberattack that can be used to understand cyber threats and the steps they use to worm their way into computer systems. Traditionally this method has been used to address the seven levels of an attack, seeking to stop the attack in the early stages. The kill chain outlines the stages in a cyber attack:
AV has evolved since then as the myriad of vendors employ hundreds of thousands of employees to keep threat signature lists up to date and as close to real time as they can. The solutions evolved to fight malware in a much more pervasive and effective way, across every possible opening, are themselves multiplying fast. Today the rate of new daily attacks is well over 300K. Corporate networks just cannot keep up.
However, so-called “Zero Day” attacks – ones without signatures, and fileless attacks, remain difficult if not impossible for AV to stop due to their very nature –– they have no history in the database, and move stealthily to target the easiest paths to penetrate the network and steal, encrypt or even erase data forever.
According to “The 2017 State of Endpoint Security Risk” report by the Ponemon Institute (November 2017), up to 77% of successful breaches today involve previously unknown malware or some type of fileless attack.
Today, AV has an alternative, and it is Artificial Intelligence (AI). AI is a set of intelligence, built and refined as an algorithm that monitors all behaviours, instantly identifies zero day attacks based on their behaviour, and then puts an immediate stop to them.
The global cyber security situation is now so dangerous that only an AI and machine learning-type approach can examine, sort, and analyse all of the data coming through any sizeable organisation in any meaningful way to make sense of it.
“The critical thing that AI is going to be useful for is tracking these unknown and new attack vectors,” says Ziften CEO Mike Hamilton.
“They make up about 77% of breaches that have been investigated to date. In my opinion this is where AI is really going to play at the macro level. AI and machine learning really stand out at helping us with ‘the signal to noise level problems.
“Machine learning is fantastic at being able to look at very small changes in very small signals in huge streams of the data to make a prediction of what's going to happen next. That's why security is such a fantastic application for AI because it can detect what humans are unable to because of the sheer size of the data sets. On the threat detection side of it, is inherently this signal-to-noise problem.”
“On the network you are typically looking for small changes in network activity, to identify anomalies that could be indicators of an attack, or whether you're looking at the endpoint level, level looking for anomalies in files, anomalies in user activities and to be able to close-in on that signal.”
"AI, due to its nature to learn the behaviour of the attacker rather than simply respond to historical attack signatures, is much more effective than traditional AV, and combined with solutions that view, monitor and assess the state of all connected assets across the enterprise, makes for the ideal way to prevent losses and fortify cybersecurity in a way that scales,” Hamilton continues.
The notable difference between traditional and the new AI-based cybersecurity systems is the difference between teaching and learning. Old antivirus software was programmed to recognise specific security threats – viruses, worms, or ransomware – based on heuristics and specific digital signatures. The software has been instructed on exactly what to look for as it scans through files - there is no intelligence to it.
A fatal problem with this approach is that it cannot handle new threats that it simply does not know about or recognise, so the anti-virus manufacturers have to constantly add new instructions to reflect the latest viruses or malware. The big anti-virus companies have hundreds of thousands of programmers updating their systems to protect against new threats.
The problem with this ‘traditional’ approach is that in recent years cyber threats have become too numerous and too sophisticated for legacy antivirus to keep up. In 2017 experts discovered over 7.4 million new malware specimens, leading some to call it the “Nightmare Year” for cybersecurity. To put that figure in perspective, it represents a 5,600% increase over the past decade.
“And recently, nation states and foreign intelligence services are putting huge amounts of treasure and sweat into creating new novel attack vectors. And that's why it is increasingly more important on the defensive side to be looking for new technologies like artificial intelligence that can predict brand new never-been-seen attacks,” Hamilton says.
“You are now able to run that AI model on your end point, and not take up more than 1% of your overall CPU power. For a lot of the neural implementations the sheer size of the data sets meant that a lot of that work happened in the cloud - in the mid 2000s we were faced with data sets of hundreds of gigabytes which were just too big to analyse.”
AI in the context of endpoint security can address not only known file-based attacks, but also unknown file-based attacks and new fileless and in-memory attack vectors. AI and machine-learning have now developed to the point where we have the available compute power to run AI at the endpoint – it now takes a fraction of the CPU power which it used to and is now affordable for all big organisations – depending on whose solution you buy!
Security threats from all angles
Not only have the type of attack vectors increased but so are the types of attackers. Some are nation-states – government sponsored-- with their own security and commercial reasons for wanting to attack companies, some are looking for purely commercial targets and some are terrorist organisations – all are growing in number and all organisations have to take them and their own security seriously. Recent advances in machine learning are enabling a more effective approach for endpoint security.
The days when you could “teach” software all of the individual threats to look for are long-gone. Engineers now fill machine learning algorithms millions and millions of examples of cyber threats, allowing the systems to “learn” how to distinguish benign from malignant.
So when hackers produce new variants of viruses and malware to avoid the signature-detection of traditional antivirus, machine-learning based endpoint security, understanding the patterns and anomalies that characterize malware, AI is rarely fooled. One recent study found that an AI-powered system identified zero-day threats correctly 98.88% of the time, compared to the 71.16% success rate of traditional AVs.
How AI specifically addresses Zero Day attacks and “In memory” attack vectors
Today, for the first time the security industry can automate the world of the threat vector. AI allows us to feed into the model millions of samples of bad code, malicious code and malware so that it can learn where the threats are most likely to occur.
Supervised machine learning techniques can teach the AI to learn what makes a file malicious or benign. And these security advances are quickly starting to make the old anti-virus approach redundant. Some AI approaches merely address file-based malware attacks.
In other words, they offer little improvement over traditional signature-based AV approaches. Other AI approaches offer multi-vector protection. But how? AI needs to address scripts, macros, and even in-memory attacks.
“As compute power produces faster and better processing it provides more machine learning frameworks for you to efficiently train large scale models,” Hamilton says.
“That's why we are really seeing the acceleration over the last five years in the allocation of AI, since we are now able to match the conceptual work in AI done in the academic community with the affordable computational side with GPUs and thought processing.”
The bottom line:
All enterprise organisations should be looking to move to endpoint protection based on an AI approach that offers multi-vector cyber attack prevention.