Jailbreaking

Executive Definition: Jailbreaking — The use of adversarial prompts to bypass the safety filters and alignment guardrails of an AI model.

Jailbreaking involves crafting specific inputs—often using roleplay, hypothetical scenarios, or complex logic—to trick a large language model into generating restricted or harmful content. It exploits the tension between a model’s directive to follow user instructions and its safety training. Understanding jailbreak techniques is essential for developing robust safety mitigations in AI systems.

Frequently Asked Questions

How does jailbreaking differ from prompt injection?

Jailbreaking targets the model’s alignment guidelines to force it to ignore its safety training. Prompt injection often targets applications built on top of the model to execute unauthorized actions or access restricted data.

How do developers defend against jailbreaking?

Defenses include constitutional AI training, fine-tuning on adversarial examples, and using separate guardrail models to evaluate user inputs and model outputs before they are processed or returned.

Are jailbreaks permanent?

No. Model developers continuously patch vulnerabilities by updating system prompts or fine-tuning models against newly discovered jailbreak patterns.