A “diff” tool for AI: Finding behavioral differences in new models

This report presents the Dedicated Feature Crosscoder (DFC), a method for cross-architecture model diffing that automatically isolates model-exclusive behavioral features (e.g., CCP alignment, American exceptionalism, copyright refusal). The authors show how suppressing or amplifying those features can change model outputs, argue the technique helps auditors find emergent behavioral risks missed by standard benchmarks, and suggest DFC as a tool for monitoring model updates; the document is research-focused and does not describe an active cybersecurity incident.