Glossary
Prompt Injection
An attack where an attacker's text is treated as instructions by an LLM.
Definition
Prompt injection is a class of attack where an adversary crafts input text designed to override the AI system's original instructions. Direct injection: a user types adversarial input. Indirect injection: an attacker plants adversarial text in a document, webpage, or other content the AI will consume.
Context
Prompt injection is the most consequential security concern in modern AI systems — analogous to SQL injection for databases. Effective defenses combine input sanitization, output validation, minimal-privilege tool use, and human confirmation for high-stakes actions. No defense is perfect; agent designs should assume adversarial inputs and limit the blast radius of any single agent action.