To clarify, the custom instructions apply to all models by default (except Research Experimental, which doesn’t support them yet), so there’s no leakage as it’s intended to be provided to the models in the first place. You may want to experiment with different prompting strategies in your custom instruction prompt to make it more subtle if you find the references too blunt. You can also try out different models like Kimi K2 which is less sycophantic in general.
