Activation steering can effectively reduce specific skills and behaviors in language models without significant negative impacts on overall performance.