What is Prompt Injection?
Prompt injection occurs when input causes an LLM to behave or respond contrary to intended instructions. It may be direct through a user request or indirect through a webpage, document, email, or retrieved record. The architectural issue is that an LLM can interpret untrusted data as an instruction. RAG and fine-tuning can improve relevance, but neither alone reliably separates data from instructions.
Prompt injection security risks should be assessed in the context of the system, data flows, identities, integrations, and decision consequences. A precise boundary helps owners evaluate the exposure without overstating what one control can achieve. See the authoritative source for the underlying reference.
What is Prompt Injection used for?
Attackers may manipulate outputs, expose information available to the application, or influence connected tools. Impact depends on the system’s authority. Separate trusted from untrusted content, minimize tool permissions, validate outputs before action, and require human approval for consequential operations. Do not claim that a system prompt or filter solves this risk.
Leaders should define accountable ownership, test relevant failure conditions, and retain evidence for decisions and change review. The response depends on the use case, authority, data sensitivity, architecture, and operational capacity rather than a universal checklist.
