Skip to content
CyberSimStart training

Theory · 9 min · AI at work: what the SOC sees

Prompt injection, direct and indirect

Tell direct from indirect prompt injection, and see why the indirect kind reaches the SOC.

A large language model (LLM) reads instructions and data in the same stream of text. It has no reliable way to tell "what my user asked" from "text that happens to contain instructions". That is the root of prompt injection, number one in the OWASP Top 10 for LLM Applications 2025 (LLM01).

Direct prompt injection is the user typing instructions that override the system’s rules ("ignore your previous instructions and..."). It mostly matters when the user is the attacker, for example against a public chatbot.

Indirect prompt injection hides the instructions in content the model reads later: a web page, an email, a document, a ticket, a code comment, a tool’s output. The victim is the user whose assistant reads it. ATT&CK v20 lists it as T1698 under Execution; MITRE ATLAS as AML.T0051.001.

No filter catches every injection. The defences that work limit what a hijacked model can do: least privilege for its tools, human approval for risky actions, and treating everything it reads as untrusted input.

Direct injection
The attacker types the malicious prompt into the model themselves.
Indirect injection
The malicious prompt sits in content the model processes on someone else’s behalf.

T1698AML.T0051.000AML.T0051.001

Your turn: sort the evidence. Direct or indirect prompt injection?

Do the interactive check Free, in the browser.

Sources

Part of