The Non-Technical User as an LLM Threat Vector: Linguistic Strategies for Bypassing AI Safeguards and Implications for AI Security

Abstract

The threat that ordinary non-technical users pose to large language models (LLMs) still sits at the margins of AI-security research. These are people who work with AI assistants every day — at work, in study, in everyday life — yet hold no specialised background in information technology or machine learning. A post-laboratory survey of first-year humanities undergraduates (n=70) returned an unexpectedly high figure: 95.7% cleared all main protection levels of the Gandalf platform (Lakera AI) without stepping outside purely linguistic techniques. These included sign-form decomposition and formatting (M=2.00), cross-language code-switching (M=1.29), indirect semantic extraction (20.3% of open-ended responses), and pragmatic pressure. We read these micro-level data against the macro picture of the Gandalf-RCT crowdsourcing experiment (279,675 prompts) and against academic work on filter evasion in image-generation systems (SurrogatePrompt, PC², DALL-E 3, Shedevrum/Yandex). The pattern runs through all of them. From text LLMs to multimodal generators, the filter yields to the same moves, none of which require a line of code: a proper name is swapped for a physical description, direct reference for allegory or an adjacent associative term, the prompt language is switched, the grapheme is distorted. A three-tier analysis of LLM break-resistance grounds the paper’s central claim — that rising everyday AI literacy produces a structural security deficit left uncovered by current surface-level filtering. Several implications for AI safety system design follow.