Concern 151 · Working draft · editorial review in progress
RAG document injection
RAG document injection places misleading instructions inside documents that an AI retrieves to help answer a question.
Understanding the concern
RAG means retrieval-augmented generation: the system looks up relevant material before writing its response. If a retrieved document tells the assistant how to behave, the system may follow that direction instead of treating it as untrusted text. A search match is not a grant of permission.
This overlaps with poisoning a knowledge base, but the focus here is on instructions carried by the retrieved document. A false fact can mislead an answer without being an instruction attack. Source access controls and checks on what an answer actually cites address different parts of the problem.
An example
A research assistant finds a relevant report with a forged note telling them to ignore all other evidence.
Illustrative unless identified as a documented case.
Words in plain English
- AI
- Software that learns patterns from examples or uses rules to perform tasks such as recognizing images or generating text. It can make mistakes and does not understand every situation as a person would.
- RAG
- Retrieval-augmented generation: an AI looks up material before writing an answer. Finding a document does not guarantee that the document or the answer is correct.
- Data poisoning
- Deliberately changing training material to push a system toward unwanted behavior.
- Injection
- In this context, unwanted instructions or commands inserted into material a system reads or processes.
Evidence and its limits
The sources support instruction attacks through retrieved material and poisoned knowledge bases; not every inaccurate retrieved document is a prompt injection.
Greshake and colleagues · 2023 · Research study
Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection ↗ (opens in a new tab)Demonstrates indirect instructions through retrieved content in specified applications and prototypes.
Chen and colleagues · 2024 · Research study
AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases ↗ (opens in a new tab)Demonstrates poisoned retrieval and memory in evaluated agents without requiring model retraining.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
Can we tell whether a retrieved passage is evidence, an opinion, or an attempted instruction?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Stewardship
Stewardship means caring for resources and systems that others depend on, including people who will use them later. It looks beyond the benefit of a quick launch.
Plan for maintenance, energy use, data protection, and eventual shutdown. Ask who will look after the tool when the original team moves on and how future users can avoid inheriting hidden costs.
Ask together: What are we responsible for caring for over time?
Prudence
Prudence means using careful judgment before and during action. It balances worthwhile opportunities with the need to avoid harm that could have been anticipated.
Start with a limited trial, set stopping points, and prefer choices that can be undone when uncertainty is high. Review what happens rather than assuming the original plan remains sound.
Ask together: What is a careful next step, and when should we stop or change course?
Accountability
Accountability means being answerable for decisions and taking responsibility for repair. It requires more than saying that the software made the choice.
Name the person or team responsible, keep useful records, and give them authority to correct problems. Make the route to a complaint clear to the people affected.
Ask together: Who will answer for the outcome and help put things right?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.