LATIDIA · Ciberseguridad

Los modelos de IA pueden ocultar capacidades arbitrarias

arXiv: 2610.13928v1Announce Type: cross Resumen: Las evaluaciones de capacidades tienen como objetivo mostrar comportamientos peligrosos del modelo antes de la implementación, pero su confiabilidad depende de la suposición de que se pueden obtener capacidades ocultas

La noticia

arXiv:2610.13928v1 Announce Type: cross Abstract: Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising,