AML.T0068

MITRE ATLAS / LLM Prompt Obfuscation

Adversaries hide or obfuscate prompt injections and retrieval content to avoid detection by humans, model guardrails, or other detection mechanisms. Encoding, invisible characters, and unusual formatting are common.

Think of it likeWriting the instruction in lemon juice, invisible until the right lamp is held over it.

In plain English

An injection does not have to be readable to work. White text, zero-width characters, base64, and image-embedded text all reach the model while a reviewer sees nothing.