Malicious Use of AI Poses a Real Cybersecurity Threat
By Erik Zouave
Could the same automated technologies cybersecurity professionals are increasingly using to protect their enterprises also fuel attacks against them? The research bears that out, according to a report my colleague Marc Bruce and I recently completed for the Swedish Defence Research Agency. The use of artificial intelligence (AI) tools to analyze data and predict outcomes has been a boon for many industries, including the cybersecurity and defense industries.
We should prepare for a future in which artificially intelligent cyberattacks become more common.
Could the same automated technologies cybersecurity professionals are increasingly using to protect their enterprises also fuel attacks against them? The research bears that out, according to a report my colleague Marc Bruce and I recently completed for the Swedish Defence Research Agency.
The use of artificial intelligence (AI) tools to analyze data and predict outcomes has been a boon for many industries, including the cybersecurity and defense industries. More and more, antivirus and cyberthreat intelligence systems are using machine learning to become more efficient. For example, both the US Defence Advanced Research Projects Agency (DARPA) and the European Defence Agency (EDA) are seeking to integrate AI technologies into their cyberdefense response capabilities.
However, AI could be a double-edged sword, with some of the industry’s most prominent thinkers warning of AI-supported cyberattacks. In this game of cat and mouse, foreseeing how AI might be used in malicious cyberattacks — and understanding its future potential — will better prepare and equip responders.
This article summarizes some of the important takeaways from our report, “Artificially Intelligent Cyberattacks.”
How Did We Get Here?
The vast and varied potential of AI misuse came into focus in a landmark report by several research institutions in 2018. The report showed the general potential for digital, physical, and social malicious uses of AI.
However, it had already been established that AI could play a prominent role in cyberattacks. The first AI-supported cyberattack, recorded in 2007, came courtesy of a dating chatbot conspicuously dubbed CyberLover, described as displaying an “unprecedented level of social engineering.” The bot relied on natural language processing (NLP) to profile targets and generate customized chat responses containing fraudulent hyperlinks, becoming notorious for its personal data thefts. It was estimated that CyberLover could establish a new relationship every third minute.
Fast forward to 2016: DARPA organized the Grand Cyber Challenge, in which machines, not humans, were the main contestants. During the contest, AI-supported solutions were used to detect, exploit, and patch vulnerabilities. It is noteworthy that the challenge not only attracted contestants from research institutions, but also the defense industrial complex.
More recently, amid increasing concerns about future AI misuse, the United Nations Institute for Disarmament Research reported on the normative and legal aspects of AI in cyber operations, reaffirming government responsibility to enact policies about use and misuse of new technology. Concomitantly, cybersecurity firm Darktrace, and IBM began looking into specific technical use cases for AI in cyberattacks.
Malicious Uses of AI in Cyberattacks
With that as our backdrop, it is vital to anchor the way forward in our response to AI misuse. Based on our extensive, peer-reviewed research of mainly experimental AI prototypes, AI’s data aggregation capabilities appear to be top-of-mind for cyberattackers who want to leverage the technology to inform their attack plans. In the short term, the absolute strongest cases for this is in the initial reconnaissance stage of cyberattacks. Through a multitude of applications, AI technologies have shown to be supremely effective at data analysis. AI cyberthreat intelligence solutions are already available, including IBM’s Watson for cybersecurity and offerings from Cylance and CrowdStrike. Hence, we can expect that AI-supported antagonists have the ability to efficiently generate intelligence on threat mitigation trends, profile targets, and generate libraries of (known) vulnerabilities at scale.
Another malicious capability to watch is the efficiency of AI in conducting repetitive tasks. As seen in the Ticketmaster incident dating back to 2010, AI tools to defeat Captchas are readily available. The experimental research on Captcha-defeating is likewise well-established. However, repetitive tasks — such as password-guessing, brute-forcing and stealing, as well as automating exploit generation — should also be considered promising grounds where prototypical testing might mature into more advanced solutions. For example, some experiments, such as password brute-forcing and password-stealing, have displayed success rates of over 50% and 90%, respectively.
Finally, deception and manipulation seem likely capability developments stemming from AI. The case for increasingly sophisticated AI-supported phishing may seem imminent when looking at a relatively old case such as CyberLover. In reality, research about AI-supported phishing, along with AI tools to bypass phishing detection, has produced mixed findings. However, this does not negate AI’s potential to statistically surpass the efficiency of human social-engineering attempts.
Already, AI-supported attacks have allegedly begun to mimic patterns of normal behavior on target networks, making them harder to detect. While network behavior analysis technology is already in use for security, research indicates this technology also could be twisted for malicious ends. Furthermore, an emergent research domain concerns the ability to attack different types of classifiers that identify patterns in data, such as spam filters that identify spam email. With new uses for NLP and other AI classifiers on the horizon, security concerns become more diverse.
We should prepare for a future in which artificially intelligent cyberattacks become more common. As mentioned, AI’s advanced data aggregation capabilities could help malicious actors make more informed choices. While these capabilities may not necessarily push a shift toward more sophisticated attacks, the potential to increase the scale of malicious activities should be of concern. Even the automation of simple attacks could aggravate trends such as data theft and fraud.
Long term, we should not discount the possibility that developments in deception and manipulation capabilities might increase and diversify the sophistication of attacks. Even experts developing AI are beginning to worry about various potential for deception, and those concerns span the fields of AI text, image, and audio generation. In a future of physical systems implementing AI, such as smart cars, an AI arms race between defenders and attackers may ultimately impact risks of physical harm.
While researchers are putting their minds to the task of securing the cyber domain with AI, one of the worrying conclusions from our research is that these efforts might not be enough to protect against the malicious use of AI. However, by anticipating what to expect from the misuse of AI, we can now begin to prepare and tailor diverse and more comprehensive countermeasures.