The article focuses on the dual-use dilemma of large language models (LLMs) in cybersecurity, highlighting how their strengths for defense also make them powerful tools for attackers. Malicious LLMs, such as WormGPT and KawaiiGPT, are specifically designed to bypass ethical constraints and safety measures, allowing cybercriminals to generate phishing emails, craft malware, and automate reconnaissance. These models are marketed in underground forums, enabling even low-skilled attackers to launch sophisticated campaigns without deep technical expertise. WormGPT emerged in July 2023 as one of the first commercialized malicious LLMs. Built on the GPT-J 6B model, it was fine-tuned with datasets containing malware code and phishing templates. This design allowed it to produce highly convincing phishing messages and generate malware snippets. The original WormGPT, despite being shut down due to negative publicity, established a template for similar models. Its successor, WormGPT 4, further evolves this concept by explicitly rejecting ethical boundaries, providing tools that make sophisticated cybercrime more accessible. The impact of these malicious LLMs is significant. They lower the barriers to entry for cybercrime, allowing individuals with minimal skills to conduct attacks that previously required more expertise. This shift enables a scale of cybercrime that can outpace traditional defenses, compressing the attack lifecycle from days or hours into mere minutes. The proliferation of these tools signals a troubling trend in the cybersecurity landscape, where offensive capabilities are becoming more mature and widely available.