Prompt Injection and Jailbreak Techniques

This chapter explores the evolving threat landscape of prompt injection and jailbreak attacks targeting generative artificial intelligence (AI) systems. Unlike traditional software vulnerabilities, these attacks exploit the model's linguistic interface, using carefully crafted language to manipulate behavior, override ethical safeguards, or trigger harmful outputs. The chapter walks through direct prompt injections, indirect attacks through retrieval-augmented systems, and conversational manipulation techniques that unfold across multiturn dialogues—revealing how attackers exploit the model's structure, context retention, and instruction-following bias. Readers will gain a comprehensive understanding of how adversaries engineer prompts to subvert safety filters, weaponize external data sources, and manipulate the model's internal state through embedding space and positional encoding vulnerabilities. The chapter illustrates how attackers iterate and optimize their techniques to reliably induce policy-violating outputs, even in well-aligned models. Key concepts such as deferred triggers, semantic boundary exploitation, and token-level perturbation are examined in depth, showing how subtle linguistic cues can bypass traditional input validation methods. For AI security professionals, this chapter offers critical insights into identifying, modeling, and mitigating prompt-based attacks. It provides structured comparisons between injection vectors, behavioral symptoms, and architectural weak points in the generative pipeline. By reframing the prompt as a high-risk interface, the chapter highlights the urgent need for layered defenses that address not only what users say—but how models interpret and respond across time, context, and retrieved knowledge.