What Security Teams Need to Know About Agents Using Abliterated Models