Since its release at the end of November, ChatGPT, an AI chatbot developed by OpenAI, has gained significant popularity, with an estimated 100 million users engaging with it. Many have been impressed by its ability to perform various tasks in a highly sophisticated and human-like manner. Following the success of ChatGPT, other tech giants such as Microsoft and Google have joined the AI bandwagon, announcing their own AI-powered language models for search and conversation.
However, not all the feedback for these advanced language models has been positive. Some concerns have been raised regarding the potential misuse of ChatGPT for nefarious purposes. For example, a professor at Wharton, University of Pennsylvania’s business school conducted an experiment where ChatGPT took an MBA exam and surprisingly scored a B/B-. This raises concerns that students could potentially use ChatGPT to cheat on exams, highlighting the potential challenges that the post-AI world may present.
Another significant concern is cybersecurity. Reports suggest that ChatGPT is already being used by malicious actors for nation-state cyberattacks. Furthermore, cybersecurity firms have demonstrated how ChatGPT can be used to create polymorphic malware, spear-phishing emails, and even share information-stealing malware code on criminal forums. These findings indicate the potential for ChatGPT to be employed in highly evasive adaptive threat (HEAT) attacks that can bypass traditional security solutions like firewalls and secure web gateways.
To gain further insights into the potential dangers posed by ChatGPT, experts asked the chatbot whether it could be misused to develop HEAT attacks. Initially, ChatGPT acknowledged the possibility, stating that if a malicious actor had access to the model and trained it on malware samples, it could generate sophisticated and difficult-to-detect malicious code or be used for phishing campaigns. However, subsequent inquiries led ChatGPT to clarify that while it could be misused for generating misleading or false information, it is not capable of generating malware on its own.
Nevertheless, concerns remain that ChatGPT could give rise to democratized cybercrime. The fear is that individuals with limited technical skills could learn to create credible social engineering tactics, phishing emails, and even write evasive malware using platforms like ChatGPT. This democratization of cybercrime has already caused catastrophic damages, as demonstrated by the booming ransomware-as-a-service industry.
Given these potential risks, organizations must prioritize enhancing their security strategies to protect against HEAT attacks. One recommended approach is to integrate isolation technology into their security measures. This involves executing all active content in an isolated, cloud-based browser rather than on users’ devices. By doing so, organizations can prevent malicious payloads from reaching their target endpoints, effectively thwarting potential attacks.
As highly evasive adaptive threats continue to rise, it is crucial for companies to adapt their security strategies accordingly. The capabilities of language models in coordinating attacks necessitate a proactive and robust approach to cybersecurity. Only by staying ahead and continuously improving their defenses can organizations effectively mitigate the risks associated with advanced AI language models like ChatGPT.
About the Author:
Brett Raybould, an EMEA Solutions Architect at Menlo Security, is passionate about security and providing solutions to organizations seeking to safeguard their critical assets. With over 15 years of experience working for various tier 1 vendors specializing in detecting inbound threats across web and email, as well as data loss prevention, Brett joined Menlo Security in 2016. He discovered how isolation technology offers a new and effective approach to address the challenges faced by detection-based systems.