Breakthroughs in large language models (LLMs) are driving an arms race between cybersecurity and social engineering scammers. Here’s how it’s set to play out in 2024.
For businesses, generative AI is both a curse and an opportunity. As enterprises race to adopt the technology, they also take on a whole new layer of cyber risk. The constant fear of missing out isn’t helping either. But it’s not just AI models themselves that cyber criminals are targeting. In a time when fakery is the new normal, they’re also using AI to create alarmingly convincing social engineering attacks or generate misinformation at scale.
While the potential of generative AI in assisting creative and analytical processes is without doubt, the risks are less clear. After all, phishing emails created using the technology are more convincing than those full of typos and grammatical errors. Profile images created in image synthesizers are increasingly hard to tell apart from the real thing. Now, we’re reaching a stage when even deepfake videos can easily fool us.
Equipped with these technologies, cyber criminals can create highly convincing personas and extend their reach through social media, email and even live audio or video calls. Admittedly, it’s still early days for generative AI in social engineering, but there’s little doubt that it will come to shape the entire cyber crime landscape in the years ahead. With that in mind, here are some of our top generative AI-driven cyber crime predictions for 2024.
Technical expertise will no longer be a barrier to entry
Crime as a service is nothing new. Cyber crime syndicates have been lurking on the dark web forums and marketplaces for years, recruiting less technically minded individuals to expand their nefarious reach.
But with the democratization of AI and data come new opportunities for non-technical threat actors to join the fray. With the help of LLMs, would-be cyber criminals need only enter a few prompts to create a compelling phishing email or a malicious script. This new generation of threat actors can now streamline the weaponization of AI.
In October 2023, IBM published a report that found the click-through rate for an AI-generated phishing simulation email was 11%, compared to 14% for humans. However, while humans emerged as the winners, the gap is closing fast as the technology advances. Given the rise of more sophisticated models, which can better mimic emotional intelligence and create personalized content, it’s highly probable that AI-created phishing content will become every bit as convincing, if not more so. That’s not even considering it can take hours to craft a convincing phishing email, whereas it only takes a few minutes using generative AI.
Routine phishing emails will no longer be easily identifiable by spelling and grammar mistakes or other obvious cues. That doesn’t mean social engineering scammers are getting smarter, but the technology available to them most certainly is.
Moreover, scammers can easily scrape data from the brands they’re trying to impersonate and then feed that data into an LLM to create phishing content that embeds the tone, voice and style of a legitimate brand. Also, given how much we tend to overshare on social media, AI-augmented data scraping is increasingly adept at taking our online personas and turning them into intimate target profiles for highly personalized attacks.
Learn more about AI cybersecurity
Custom open-source model training will advance cyber crime
Most of the popular generative AI models are closed-source and have robust safety barriers built in. ChatGPT won’t knowingly generate a phishing email, and Midjourney won’t knowingly generate a compromising image that could be used for blackmail. That said, even the most stringently monitored and secured platforms can be abused. For example, people have been trying to jailbreak ChatGPT ever since it came out, using the so-called DAN (do anything now) prompts to get it to act without filters or restrictions.
We’re now in the midst of an arms race between model developers and those who seek to take them beyond their predefined limits. For the most part, this comes down to curiosity and experimentation, including among cybersecurity professionals who want to know what they’re up against.
The bigger risk lies in the development of open-source models, such as Stable Diffusion for image synthesis or GPT4ALL for text generation. Open-source LLMs can be customized, expanded and unleashed from any arbitrary constraints. Moreover, these models can run on any desktop computer equipped with a sufficiently powerful graphics card, far away from the watchful eyes of the cloud. While custom and open-source models typically require a degree of technical expertise, especially when it comes to training them, they’re certainly not restricted to experts in malware development or data science.
Cyber crime syndicates are already developing their own custom models and selling them via the dark web. WormGPT and FraudGPT are two such examples of chatbots used for developing malware or carrying out hacking attacks. And, just like the mainstream models, they’re under constant development and refinement.
Live deepfake scams will become a serious threat
In February 2024, CNN reported that a finance worker at a multinational firm was scammed into paying out $25 million to fraudsters. This wasn’t the sort of phishing email that most of us are familiar with. Rather, it was a deepfake video in which the scammer used generative AI to create an avatar that convincingly impersonated the company’s chief financial officer during a live conference call.
One could be forgiven for thinking that such an attack sounds like something straight out of a dystopian science fiction scenario. After all, what seemed outlandish just a few years ago is now on its way to becoming the number-one attack vector for sophisticated and highly targeted social engineering attacks.
A recent report found that 2023 alone saw a 3,000% increase in deepfake fraud attempts, and there’s no reason to believe this trend won’t continue through 2024 and beyond. After all, face-swapping technology is now readily available, and like every other form of generative AI, it’s advancing at a pace that’s near impossible for lawmakers and infosec professionals to keep up with.
The only thing holding deepfake video scams back is the substantial computing requirements involved, particularly for scams carried out in real time. A more immediate concern, especially in the foreseeable future, is the ability of generative AI to mimic voices and writing styles. For example, Microsoft’s VALL-E can create a convincing clone of someone’s voice from a three-second audio recording. Even handwriting isn’t immune from deepfakes.
How can organizations and individuals protect themselves?
Like almost any disruptive innovation, generative AI can be a force for good or bad. The only viable way for infosec professionals to keep up is to incorporate AI into their threat detection and mitigation processes. AI solutions also provide the tools needed to improve the speed, accuracy and efficiency of security teams. Generative AI specifically can assist infosec teams in operations like malware analysis, phishing detection and prevention and threat simulation and training.
The most effective way to keep ahead of cyber criminals is to think like cyber criminals, hence the value of red-teaming and offensive security. By using a similar set of tools and processes to those used by threat actors, infosec professionals are better equipped to stay a step ahead.
By understanding how the technology works and how malicious actors are using it, businesses can also train their employees more effectively to detect synthetic media. In an era when it’s easier than ever to impersonate and deceive, it has never been more important to defend reality against the rising tide of fakery.
If you’d like to learn more about cybersecurity in the era of generative AI and how AI can enhance the abilities of your security teams, read IBM’s in-depth guide.
The post Social engineering in the era of generative AI: Predictions for 2024 appeared first on Security Intelligence.