← AI security in plain English

AML.T0065

MITRE ATLAS / LLM Prompt Crafting

Adversaries use what they have learned about a target generative AI system to write prompts that get past its defenses and cause malicious instructions to run. Crafting is usually iterative, refined against observed refusals.

Think of it likeRehearsing exactly how to word a request to a strict receptionist until one phrasing gets you waved through.

In plain English

Prompt attacks are rarely one lucky sentence. The adversary probes what gets refused, learns the shape of the guardrail, and rewrites until a phrasing slips past it.