LLMs and how Attackers Can Use them
Constantin Nilă, Georgiana Diana Marin, IPSOS Interactive Services SRL, Victor-Valeriu Patriciu · Journal of Military Technology · 2023
Over the past year, Large Language Models (LLMs) have shaped the field of artificial intelligence (AI).Applications and tools such as ChatGPT, Gemini, Deepinfra, and the array of options from HuggingFace have shown significant advancements in natural language processing (NLP).These models can generate human-like text, understand complex instructions, and produce creative content.LLMs can simplify concepts into more understandable explanations, democratizing access to information and allowing users to interact more naturally and conversationally.Our paper delves into the ethical quandaries associated with LLMs, highlighting concerns over bias, privacy breaches, and the risk of abuse and disinformation.We examine the ethical implications derived from the use of LLMs while also touching on the main privacy concerns demonstrated throughout three main attack types.Furthermore, our analysis concentrates on how online services, which render these technologies consumer accessible, can be exploited by potential threat actors.We explore their use as attack vectors in strategies such as watering holes, highlighting the vulnerabilities inherent in exposing advanced technologies without proper testing.Emphasizing a constructive perspective, we recognize LLMs as a transformative force in technology, urging an approach that welcomes this innovation while diligently mitigating potential risks.