Online information spaces are where people encounter and judge institutional communication, making federal research agency press releases attractive targets for large-scale persuasive interference. We study this cyber social threat by generating 134,136 LLM-based persuasion attacks against 972 press releases from ten agencies using GPT-4, Gemma 2, and Llama 3.1 across 23 SemEval 2023 Task 3 techniques. We then characterize the moral framing of these attacks. Across models, we find distinct and technique-dependent moral signatures. GPT-4 most often centers Care, with Authority and Loyalty commonly secondary. Gemma 2 places greater emphasis on Care and Authority, whereas Llama 3.1 leans more strongly toward Loyalty alongside Care. These results show that model choice materially reshapes the moral framing of generated persuasion attacks. Identifying such signatures is an important step toward proactive defense against narrative manipulation in online ecosystems.

PDF