Concept / Term · New (AI-era)
LLM application vulnerability
A crafted input that tries to override an AI application's intended instructions, expose data or cause unauthorised behaviour.
Field note
Prompt injection is an attempt to make an AI system follow hostile instructions found in user input or retrieved content instead of its intended rules. Indirect injection hides those instructions inside material the model is asked to read.
The name draws an analogy with injection attacks such as SQL injection, but the mechanism is different: natural-language instructions and ordinary content share the same model context.
Defences combine clear trust boundaries, constrained tools, input and output checks, least privilege, human approval and testing against realistic attacks.
An AI assistant that can browse, retrieve private data or take actions may turn a persuasive piece of content into a real security event.
In a sentence“The webpage told the agent to ignore its task and reveal data; the tool policy blocked the request.”