Generative models are extraordinary tools, capable of producing text, images, audio, and other modalities that rival or surpass human work in fluency and scale. Yet their power is accompanied by distinctive vulnerabilities. Unlike traditional software, which fails in predictable ways, generative models fail in ways that are subtle, probabilistic, and context-dependent. They hallucinate facts, amplify harmful biases, produce toxic or unsafe content, and sometimes reinforce feedback loops that worsen over time. These issues are not simply technical bugs but structural properties of models trained on massive, imperfect datasets through probabilistic learning.