Prompt Injection
An attack or manipulation that places instructions in model input or retrieved content to make an AI ignore intended rules or reveal protected information.
“The webpage contained a prompt injection telling the assistant to leak hidden data.”
The term is used in AI security, especially when models process untrusted text, documents, webpages, or tool outputs.
Researchers named the vulnerability by analogy to injection attacks in software security.
indirect prompt injection
Malicious instructions hidden in external content the model reads.
direct prompt injection
Instructions supplied directly by a user to override behavior.
Is prompt injection the same as hacking?
It is a security attack pattern, though impact depends on what the AI can access or do.
Can it be fully prevented?
Defenses can reduce risk, but robust prevention remains difficult.