AML.T0056

MITRE ATLAS / Extract LLM System Prompt

Adversaries attempt to recover a language model's system prompt, the hidden instructions that define its role, constraints, and available tools.

Think of it likeGetting the actor to recite the director's private notes instead of their lines.

In plain English

System prompts often contain business logic, internal names, and the exact wording of the guardrails. Extracting one turns guesswork into a precise map of what to attack next.