WEAPONIZING LARGE LANGUAGE MODELS: AUTOMATED PHISHING, SOCIAL ENGINEERING, AND MALWARE GENERATION

Large language models (LLMs) have essentially changed the game in natural-language generation. So far, openmodels and a few powerful commercial assistants have been able to write emails, code, and answer questionswith a human-level fluency. Sadly, the factors that bring about the great success of LLMs - contextualawareness, coherence, and ease of use - are the very factors that are being exploited by adversaries.Phishing as the most common method of social engineering is still going strong: around 96 % of attacks aredelivered by emails, and lately, several studies have inferred that 82 % of phishing emails are written with thehelp of LLMs . Results of evaluations conducted in mid-2025 showed that messages of spear-phishing generatedby LLMs had click-through rates of more than 30 %, thus, they were more successful than those created byhumans . Simultaneously, the underground communities are jailbreaking open models and launching “darkLLM” services like WormGPT, FraudGPT, and DarkBERT, which get rid of safety guardrails and provideready-made prompts for business-email compromise, code obfuscation, and malware generation, respectively, astheir main functionalities. Besides that, the latest studies reveal that autonomous prompt-injection worms can goon RAG systems, and malware examples like PromptLock and LameHug can have LLMs as a part of theirpayload that they directly embed into the payload.This article goes through how LLMs are exploited for phishing, social engineering, and malware, explains therecent case studies, and talks about the ethics of building the resilient AI systems. We, by integrating academicresearch, industry reports, and dark-web observations, give a well-structured overview of the currentcapabilities, point out the gaps in the defense, and suggest the responsible AI development as a way forward.