I'd be interested to know whether it's possible to contaminate public data to make trained models do this.

Something like a commonly repeating conditional in the public domain embedding itself in the model's logic.