Representation as a Bottleneck for Mechanistic Interpretability: The Manifestation Unit Protocol

This arXiv feed collects ML/AI research with multiple security-relevant implications. Notable items: Manifestation Units proposes a typed schema for mechanistic interpretability that makes component-level analyses more structured and queryable (eases reuse for both audits and adversarial model probing). SemiScope analyzes semi-supervised pipelines for security classification (phishing/malware), finding most gains come from classifier hyperparameter optimization rather than complex SSL interactions—practical guidance for defenders. FoGS presents a filtered mixture-of-generators for fully-synth­