Jailbreak suffixes are adversarial sequences of characters or tokens appended to user prompts to bypass the safety alignment, guardrails, and content moderation filters of large language models. Typically crafted using automated optimization algorithms such as gradient-based search methods, these token sequences are mathematically constructed to steer a model into producing restricted, harmful, or policy-violating responses. Although jailbreak suffixes frequently appear nonsensical or ungrammatical to human readers, they effectively manipulate the internal token probability distributions of a model, compelling it to output compliant completions to prohibited requests and circumvent its safety training.
