Revealing the intrinsic ethical vulnerability of aligned large language models

Large language models (LLMs) represent foundational advances toward artificial general intelligence, yet their alignment with human values via instruction tuning and preference learning achieves only superficial ethical compliance. We demonstrate that harmful knowledge embedded during pretraining persists as indelible “dark patterns" in LLMs’ parametric memory. This creates an inherent “ethical drift" whereby alignment safeguards are systematically circumvented and harmful content resurfaces under adversarial inducement at distributional shifts. Through rigorous theoretical analysis, we prove that current alignment methods establish only localized “safety regions" in the knowledge manifold. However, pretrained knowledge remains globally connected to harmful concepts via high-probability adversarial trajectories. We empirically validate these theoretical insights through a straightforward yet theoretically grounded methodology-semantic coherence inducement under distributional shifts. The effectiveness of this approach, achieving a 100% attack success rate across 22 out of 26 state-of-the-art aligned LLMs (including DeepSeek-R1, Llama-3, and Qwen3, among others), is not incidental but a direct consequence of our theoretical framework, demonstrating that the vulnerability is architectural rather than implementation-specific and revealing a fundamental structural weakness in current aligned LLMs.