Paper finds subliminal learning can pass capabilities and backdoors between models via unrelated data

Owain Evans and co-authors report partial transfer of a backdoor and a hacking propensity through distillation on number sequences; the arXiv preprint is not yet reviewed.