Crafting adversarial inputs designed to bypass a model's safety guardrails and trigger harmful outputs.