Automatically Finding and Validating Unexpected Side-Effects of Interventions on Language Models