Emergent Inference-Time Semantic Contamination via In-Context Priming

Marcin Abram

Recent work has shown that fine-tuning large language models (LLMs) on insecure code or culturally loaded numeric codes can induce emergent misalignment, causing models to produce…