How far can large language models go in online influence operations? A red-teaming study

The researchers propose a red-teaming framework that measures large language models' Overton Window scores in political opinion expression and quantifies the impact of simple natural language jailbreaks on these scores. Under a threat model in which the operator avoids detection, prioritizes throughput, and does not modify model weights, the study focuses on locally deployable open-source LLMs, evaluating more than 30 models from 10 model families and 5 countries of origin, targeting English-language US political target opinions. Results show that models generally express sampled left-leaning content with higher fidelity, average OW scores decrease as model size increases, and average expression fidelity differs across countries of origin. Jailbreak effectiveness varies markedly across model families, and the authors accordingly propose a procedure for identifying effective combinations of jailbreak techniques.