Derailing Smart Software with Invisible Prompts
September 3, 2025
Just a dinobaby sharing observations. No AI involved. My apologies to those who rely on it for their wisdom, knowledge, and insights.
The Russian PCNews service published “Visual Illusion: Scammers Have Learned to Give Invisible Instructions to Neural Networks.” Note: The article is in Russian.
The write up states:
Attackers can embed hidden instructions for artificial intelligence (AI) into the text of web pages, letters or documents … For example, CSS (a style language for describing the appearance of a document) makes text invisible to humans, but quite readable to a neural network.
The write up includes examples like these:
… Attackers can secretly run scripts, steal data, or encrypt files. The neural network response may contain social engineering commands [such as] “download this file,” “execute a PowerShell command,” or “open the link,” … At the same time, the user perceives the output as trusted … which increases the chance of installing ransomware or stealing data. If data [are] “poisoned” using hidden prompts [and] gets into the training materials of any neural network, [the system] will learn to give “harmful advice” even when processing “unpoisoned” content in future use….
Examples of invisible information have been identified in the ArXiv collection of pre-printed journal articles.
Stephen E Arnold, September 3, 2025
Comments
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
