Safety & Alignment
Prompt Injection
Attack where untrusted input overrides the developer's intended model instructions.
Definition
Prompt injection slips malicious instructions into user inputs, retrieved documents, or tool outputs — instructing the model to ignore its system prompt, leak data, or take harmful actions. Defences include input filtering, output validation and least-privilege tool design; no defence is bullet-proof.
Common use cases
- Threat modelling
- Red-team testing
- Input sanitisation