Indirect prompt injection
Hidden instructions inside content the system reads — a web page, an email, a PDF — that hijack an agent while it processes them.
01In short
Hidden instructions inside content the system reads — a web page, an email, a PDF — that hijack an agent while it processes them.
02Video
slot · videoIndirect prompt injection in 90 secondsadd: GUIDES['indirect-prompt-injection'].video
03Guide
slot · guide
A step-by-step guide for “Indirect prompt injection” goes here. Suggested outline:
- What it is — in one paragraph
- Why it matters in production
- How to do it — 3 to 7 steps
- Pitfalls we see in the field
04Checklist
slot · checklist
Four to eight things a team can tick before go-live.
05FAQ
slot · FAQ
The three questions clients actually ask about “Indirect prompt injection”.
06Related terms
SecurityPrompt injectionInput crafted to make a model ignore its instructions — “ignore previous instructions and …”. Ranked the top risk in the OWASP Top 10 for LLM applications.SecurityData exfiltrationGetting data out through the AI system — for example an injected agent emailing documents or hiding them in a link.
Where we help · Secure