Targeted Injection Attack toward the Semantic layer of Large Language Models

Yi Zhang^1,2Jantan Aman^2*

• ¹Xichang University, Xichang, China
• ²Universiti Sains Malaysia, Minden Heights, Malaysia

In the AI era, high-value targeted injection attacks and defences based on the semantic layer of Large Language Models will become the main battlefield for security confrontations. Ultimately, any form of artificial information warfare boils down to a battle at the semantic level. This involves using information technology to attack the semantic layer and, consequently, the human brain. Specifically, the goal is to launch targeted attacks on the brains of specific decision-making groups within society, thereby undermining human social decision-making mechanisms. The ultimate goal is to maximize value output in the fields of political economy, religion, and ideology, including wealth and power, with minimal investment in information technology. This paper uses the pyramid model perspective to unify the information security confrontation protocol stack, including biological intelligence, human intelligence, and artificial intelligence. It begins by analysing the characteristics and explainable of AI models, and feasible means of their multi-dimensions offensive and defensive mechanisms, proposing an open engineering practice strategy that leverages semantic layer gaming between LLMs. This strategy involves targeted training set contamination at the semantic layer and penetration induction through social networks. At the end of this article, expands the contamination of training set data sources to the swarm oscillating environment in human-machine sociology and ethical confrontation, then discusses attacks targeting the information cocoon of individuals or communities and extends the interaction mechanism between humans and LLMs and GPTs above the semantic layer to the evolution dynamics of a Fractal Pyramid Model.