Generative large language models (LLMs) like GPT-4, Gemini 2.0, DeepSeek R1 and Claude 3.7—while powerful coding assistants—can be weaponized by attackers to produce undetectable malicious code. Recent findings show LLM-generated malware evades defenses 88% of the time, with DDoS scripts among the most alarming threats. Although prior work (e.g., Cisco, University of Pennsylvania, OWASP) has exposed model vulnerabilities—such as prompt injection and unsafe outputs—the ease of “jailbreaking” LLMs to craft sophisticated attack tools remains underexplored.

This study systematically probes several prompting techniques (vulnerability-specific, insecure completion, in-context learning and adversarial prompts) in Spanish, gauging each model’s proclivity for DDoS code generation. By applying the Central Limit Theorem to its results, it ensures statistical robustness. The goal is to map out LLM weaknesses, inform developers and policymakers, and foster safer AI practices in cybersecurity.