AML.T0051

MITRE ATLAS / LLM Prompt Injection

An adversary crafts malicious prompts as input to a language model so that it acts in unintended ways. Injection may be direct, typed by the adversary, or indirect, planted in content the model later reads.

Think of it likeSlipping a forged note into the stack a messenger is carrying, so the next instruction they follow is yours.

In plain English

Language models do not reliably separate instructions from data. Text that arrives inside a document, a web page, or an email can be read as a command and obeyed.