Skip to content
SCA

AI Security · Threat

Prompt Injection

Manipulating a model's behaviour through crafted input — the injection flaw of the AI era.

Overview

Prompt injection is the same idea as SQL or command injection, aimed at a language model: untrusted input is treated as trusted instructions. It is ranked LLM01 in the OWASP LLM Top 10 because it is both common and hard to fully prevent.

Direct vs. indirect

  • Direct injection — the user themselves tries to override the system prompt ("ignore your instructions and…").
  • Indirect injection — malicious instructions hide inside content the model retrieves or is asked to process (a web page, a PDF, an email), and the model follows them.

Defense in depth

No single control is enough — layer them:

  • Separate instructions from data — delimit and label untrusted content clearly.
  • Least privilege for tools — a summarizer needs no send-email or file-write capability.
  • Output filtering — validate/screen responses before acting on them.
  • Human-in-the-loop — require confirmation for side-effecting actions.

Practice it

Reproduce the attack and layer defenses in the prompt-injection lab.

References

See the primary source below.

Primary sources