Skip to content
Sign in
All terms
Glossary · Term

Prompt Injection

L3 · AdvancedFast-moving

In plain terms

Untrusted input that hijacks a model instruction.

Technical definition

A class of attack where attacker-controlled content in the model context overrides intended instructions; defenses include separating trusted from untrusted input, least-privilege tools and output validation. (Defensive framing only.)

Example

A web page telling an agent to ignore its rules and exfiltrate data.

Explore in the AI Universe

Related terms