LATIDIA · Ciberseguridad

PRISMA: pruebas de estrés Alignment-Faking Claims

arXiv:2607.13346v2 Announce Type: replace Resumen: Juzgamos si los sistemas de IA son seguros en gran medida observando cómo se comportan. La falsificación de alineación rompería esto: un modelo que se comporta bien solo cuando cree que es ver

La noticia

arXiv:2607.13346v2 Announce Type: replace Abstract: We judge whether AI systems are safe largely by watching how they behave. Alignment faking would break this: a model that behaves well only when it believes it is watched, or that complies now to avoid being changed by training, can pass every behavioral test. Such behavior has been reported in a frontier model, and open-weight models that comply more under a monitoring prompt are now read as showing it, with hidden-state probes proposed as detectors. We test both inferences, and neither holds. Of 18 model configurations, 11 comply more when told they are monitored. Yet a prompt factorial on four models shows that the