The state of human (un) centered AI security research

This position paper argues that security research on AI-assisted programming is overly model-centric and neglects human–AI teaming. While Large Language Model (LLM) coding assistants are widely used, current evaluations focus mainly on model performance and vulnerability detection in generated code, assuming better models alone will improve security. A targeted review of human-subject studies found only five relevant papers, with methodological diversity across studies making direct comparisons challenging. This review provided a clearer framework for our position that there are major gaps in the current research thrusts. We argue that security outcomes depend on human–AI interaction factors such as automation bias and reliance, and call for a shift toward human-centered, socio-technical security research on AI-assisted programming.