AML.T0054

MITRE ATLAS / LLM Jailbreak

Adversaries induce a language model to ignore, circumvent, or override its safety and alignment behavior in order to elicit output the model was intended to withhold. Techniques include role play, encoding, and multi-turn pressure.

Think of it likeTalking a rule-following clerk into an exception, one small concession at a time, until the rule no longer applies.

In plain English

Guardrails are behavior, not walls. Enough framing, indirection, and persistence will usually convince a model that this particular request is the exception.