Definition
An attack which exploits the concatenation of untrusted input with a prompt constructed by a higher-trust party such as the application designer
Quoted definition: NIST AI 100-2e2025, prompt injection, glossary entry, sourced to NIST AI 100-2e2025.
Notes to entry
Note 1 to entry: NIST defines the general attack. In the course, an instruction embedded in an email or other source material is content to inspect, not authority for the assistant to act.
Note 2 to entry: Separating untrusted content from instructions and withholding unnecessary write tools can reduce exposure. A written refusal alone is not a technical access control.
Source basis
The definition is quoted from the named source. Notes and concept links are AidGPT additions. Other sources provide context and do not endorse this page.
- NIST AI 100-2e2025, prompt injection (National standards body terminology)
- AidGPT guided course, Sessions 1 to 5 (AidGPT course material)