Prompt injection is an attack on applications built on language models in which injected instructions cause the model to disregard its original rules. In the direct variant, the attacker writes the instruction into the input. In the indirect variant, the instruction is hidden in content that the model processes later, such as a web page, an email or an uploaded document. The OWASP Top 10 for LLM Applications ranks prompt injection first.
The attack becomes dangerous as soon as the model may use tools, read data or trigger actions. No complete defence is known so far. What counts is therefore limiting the possible damage: least privilege, separation of instructions and external content, input and output filters and human approval for consequential steps. More under MU-09 Manipulation of the inputs.