Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation

arXiv·high signal

Fine-tuning LLMs for security classification can introduce token-level evasion vulnerabilities that standard held-out evaluation on the same data distribution completely misses. The work shows models inherit circuits while learning semantics, creating blind spots that adversaries can exploit. A direct warning for builders shipping fine-tuned guardrail or moderation classifiers and trusting in-distribution accuracy.