Neurosama is a fun example but we dont really know the sauce vedal coocked up.

When i say proven i mean 32 page research paper specifically looking into it.

arxiv.org/abs/2407.12831

They found that even a model trained specifically on honesty will lie if it has an incentive.