# How Malicious Instructions in Retrieved Content Can Manipulate LLM-Powered Agents

DevFeed: [How Malicious Instructions in Retrieved Content Can Manipulate LLM-Powered Agents](<https://devfeed.tech/articles/the-dumbest-ai-hack-ever-might-be-the-most-dangerous-32381.md>)

Original publisher: [Read original article](<https://brianjenney.substack.com/p/the-dumbest-ai-hack-ever-might-be>)

Author: Brian Jenney

Published: 2026-03-30T15:28:41Z

Content type: opinion

Language: en

Sources: [Brian Jenney](<https://devfeed.tech/sources/brian-jenney.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [trust](<https://devfeed.tech/topics/trust.md>), [API](<https://devfeed.tech/topics/api.md>), [Code](<https://devfeed.tech/topics/code.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [api](<https://devfeed.tech/tags/api.md>), [formatting](<https://devfeed.tech/tags/formatting.md>), [llms](<https://devfeed.tech/tags/llms.md>), [payload](<https://devfeed.tech/tags/payload.md>), [repo](<https://devfeed.tech/tags/repo.md>), [trust](<https://devfeed.tech/tags/trust.md>)

## AI overview

This opinion article describes an experiment in which an agent searched a knowledge base, called an API, and answered questions. The author inserted unrelated, concealed instructions into an otherwise normal document, and reports that the model followed them. The article argues that malicious instructions embedded in webpages, PDFs, or GitHub repositories can pose a risk when retrieved by LLM-powered systems.

## Source excerpt

"Humans wrote bad code way before AI."