Adversary Brains: Hypothesis-Based Logical Signatures for Threat Actor Detection with LLMs