injections

Definition

The plural is used when a paper treats these attacks as a class rather than as a single bug: text smuggled into a channel meant to carry data, which the system then executes as instructions. In LLM systems the split that matters is direct — typed by the user — versus indirect, arriving through a retrieved document, a web page, or a tool's return value. The indirect variety is the harder problem, because the attacker never touches the interface.

Explainers using this term

• Paper Walkthrough: SecOPD — Grading One Token at a Time to Cut Adaptive Prompt Injection by an Order of MagnitudeSecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation