A new preprint on arXiv examines a security risk introduced by a class of AI models called System One models, specifically those built on the Jev framework. These models work by converting open-ended semantic contexts into structured judgments that then directly control what a program does. That mapping, the authors argue, creates a new attack surface.

The paper, titled "Nudge the Judge," reports the discovery of what it calls "program decision hijacking" in real-world Jev applications. While the abstract does not detail specific attack methods, it frames the issue as a fundamental consequence of the paradigm: when semantic understanding becomes the basis for program execution, influencing that understanding can influence the program itself.

Because this is a single preprint, the findings have not yet been peer-reviewed. Still, the work highlights a growing concern as AI models move from advisory roles into direct decision-making positions within software systems.