Safety and MLOps
Jailbreak
A jailbreak is an input designed to bypass a model’s behavioral safeguards.
How it works
Attackers use reframing, encoding, or multi-turn pressure to elicit disallowed behavior. Defense requires adversarial evaluation and layered controls outside the model.
A concrete example
Illustrative: a red-team set contains 500 attack prompts across 10 strategies.
Do not confuse it with Prompt injection
a jailbreak targets model safeguards directly; prompt injection can exploit untrusted external content.