LLMs Generate Harmful Content Using a Distinct, Unified Mechanism Despite Alignment Training

arXiv·medium signal

Orgad, Wei, and Zheng discover that LLMs use a single, unified internal mechanism to generate harmful content across categories — meaning jailbreaks exploit one shared pathway rather than separate vulnerabilities per harm type. This explains why alignment safeguards remain brittle: patching individual failure modes doesn't address the underlying shared mechanism. Mechanistic insight that could inform more robust alignment approaches by targeting the unified pathway rather than playing whack-a-mole with individual jailbreaks.