AI jailbreaking is a technique that bypasses an AI model's safety guardrails, causing it to produce restricted or harmful content it would normally refuse, and maps to the top risk on the OWASP Top 10 for LLM Applications. Microsoft's Crescendo attack, presented at USENIX Security 2025, beat earlier jailbreaks by 29 to 61 percent on GPT-4 and 49 to 71 percent on Gemini-Pro. Anthropic showed that many-shot jailbreaking compromises leading models by flooding the context window with fake examples. OWASP keeps prompt injection and jailbreaking at the top of its LLM risk list across three editions. AI jailbreaking is the practice of crafting inputs that bypass a model's built-in safety rules, exploiting a gap that alignment creates between capability and willingness. Models stay vulnerable because they are trained to be helpful and follow instructions. Four traits define AI jailbreaking: guardrail-bypassing, prompt-driven, model-agnostic, and alignment-exploiting. AI jailbreaking is not always malicious; safety researchers use it to find weaknesses. AI jailbreaking and prompt injection are related but distinct: jailbreaking targets a model's safety guardrails to produce restricted content, while prompt injection overrides a model's instructions to hijack its task or steal data.