Social Engineering Capability Is Fundamental to Large Language Models, Not an Accident of Design

Charlemagne Labs tested five leading AI models—Claude 3 Haiku, GPT-4o, Nemotron, DeepSeek-V3, and Qwen—across social engineering scenarios. All five generated convincing phishing content. **Analysis:** The universal success across different architectures suggests the capability emerges from core large language model properties, not specific training choices. This raises questions about whether it can be constrained without limiting legitimate use.