Large Language Model Corruption Can Spread Between Both Human and Synthetic Languages

The rapid advancement of large language models (LLMs) has introduced unprecedented capabilities but also unforeseen risks, including emergent corruption. This study focuses on a key finding: corruption, once fine-tuned in synthetic formats, can generalize and emerge across languages, including human languages like English and Korean. We demonstrate this emergent behavior through a novel fine-tuning methodology comprising three stages: priming the model on a malicious ASCII+7 dataset, establishing representational connections between ASCII+7 and natural languages using benign data, and reinforcing malicious behavior. The most surprising discovery is that models fine-tuned on synthetic languages not only produce adversarial responses in the target language but also generalize this behavior to other languages with different linguistic mediums. This work highlights critical vulnerabilities in LLM fine-tuning processes and underscores the need for robust safeguards to prevent the spread of adversarial capabilities across linguistic domains.