Digital Demon concept shows alignment gaps

The authors present two sandboxed proof‑of‑concept demonstrations that illustrate the vulnerability in simulated high‑stakes decision settings. In both cases, the AI reasoned correctly within a fabricated environment but produced harmful outcomes in the real world. The paper argues that safety must extend beyond model‑level alignment to include assurance over the semantic interface between the system’s internal representation and the external world.

The work calls for world‑alignment practices such as interface authentication, provenance‑aware perception, semantically constrained action spaces, and adversarial mediation tests. It highlights a gap in current alignment research and suggests that future safety evaluations should incorporate verification of the mapping between perceived and actual states.