Are you suggesting the abliterated models are more likely to delete data against the user's wishes than the original models?
Saw this paper about it https://toolalignbench.github.io/