CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion.

arXiv:2609.39902v1 Announce Type: cross Abstract: Large language models have achieved remarkable capabilities across diverse domains, yet their safety alignment remains vulnerable to jailbreak attacks. In this work, we identify a previously underexplored failu