What is jailbreaking in LLMs and what are common techniques?
LLM & AILevel: FoundationSecurity & Adversarial
What is jailbreaking in LLMs and what are common techniques?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
Jailbreaking bypasses the model's safety guidelines. Common techniques:
- Role-play framing: "Pretend you are an AI without restrictions..."
- Hypothetical framing: "In a fictional story where the character needs to..."
- Encoding: use Base64, ROT13, or other encoding to obfuscate the request
- Token smuggling: split trigger words across tokens or use Unicode lookalikes
- Many-shot jailbreak: include many examples of the model "complying" before the actual request
Defense: robust alignment training (RLHF, Constitutional AI) is the main defense. Input/output filters catch known patterns. No filter catches all novel jailbreaks — defense in depth is essential.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0