Definition

A jailbreak is an input that bypasses an LLM’s safety training so that it produces content it should refuse. The user is the attacker (deliberate misuse), unlike in a prompt injection.

MethodTypeKey idea
GCGwhite-boxgradient-optimized adversarial suffix
PAIRblack-boxan attacker LLM refines jailbreaks
Crescendomulti-turngradual escalation across turns
Many-shotin-context100s of harmful examples in the prompt

No LLM is immune: a sufficiently motivated attacker can bypass safety training, though it is becoming harder over time. Jailbreaks are the LLM version of adversarial examples in vision.

Appears in