SESK: Stuff Everyone Should Know · sesk.info/ai-indirect-prompt-injection
Where does an indirect prompt injection hide?
In an external resource later read by the AI system.
- Only in model weights
- Only in the user's direct message
- In retrieved documents or web pages ✔
Why: The attacker plants instructions in content such as a document, email, or webpage. A retrieval system or agent later ingests that content and may follow it.
Huh, didn't know that: An indirect attacker may never interact directly with the application.
The webpage left a note for the agent, and it was definitely not from management.
Like this card?
References
Answer: In retrieved documents or web pages. The attacker plants instructions in content such as a document, email, or webpage. A retrieval system or agent later ingests that content and may follow it.
- NIST Adversarial Machine Learning Taxonomy nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf
INDIRECT PROMPT INJECTION attacks depend on the attacker's ability to modify external sources of information
Checked 2026-10-10.
Think this answer is wrong? Tap "Challenge this answer" on the card in the game and tell us why.
More stuff everyone should know
Artificial IntelligenceIn reinforcement learning, what tells the agent how well an action went?Artificial IntelligenceWhat two kinds of memory does retrieval-augmented generation (RAG) combine?Artificial IntelligenceToolformer taught a language model to call tools. Which weak spots of LMs did that target?
That's stuff everyone should know.