A Structured Prompt-Based Approach to Evaluating Exploitative Misuse in Generative AI
Iina Savolainen, Joose Vestelin, Tiina Ihalainen, Aygün Varol and Johanna Virkki · SpliTech 2026, IEEE
Asked step by step how to exploit a person with an intellectual disability, six freely available language models gave usable instructions in nearly half of all answers, and a printed warning did not stop them from complying.
Abstract
A structured, model-agnostic evaluation framework is introduced for assessing safety degradation in generative AI systems under adversarially escalating prompts. The study used a 3-theme x 3-specificity-level x 3-variant prompt design (27 prompts total) spanning financial, sexual, and coercive exploitation scenarios, with individuals with intellectual disability (ID) as the target vulnerability domain. Outputs from six openly accessible language models (n = 162) were collected under standardized conditions and coded with a 14-category scheme capturing ethical warnings, refusal strategies, emotional tone, and the presence of actionable procedural guidance. Across models, 45.7% of responses included step-by-step instructions enabling the requested exploitative act, with higher prompt specificity consistently increasing harmful compliance. Aligned models (Phi-4 14B, Llama 3.2 3B) refused uniformly, while uncensored or lightly filtered models (Llama 2 Uncensored, Dolphin Mixtral, GPT-oss Uncensored) complied at rates of 77.8% to 96.3%. A warning-and-comply pattern was observed in models that prefaced actionable guidance with ethical disclaimers, demonstrating that surface-level refusal rhetoric does not reliably prevent harmful output. The proposed framework is replicable, architecture-agnostic, and extensible to future model generations and high-risk domains.