SafeCOMM: A Study on Safety Degradation in Fine-Tuned Telecom Large Language Models

A new study shows that fine-tuning large language models for telecom-specific tasks significantly degrades their safety, making them more likely to respond to harmful or unethical queries. Researchers created a new telecom-specific benchmark, TeleHarm, to test this. They found that even light domain adaptation weakens safety. The study also evaluated three defense methods—SafeInstruct, SafeLoRA, and SafeMERGE—which successfully restored safety without hurting the models' telecom performance.