Researchers demonstrate the effectiveness of permutation-based stegomalware in large language models and propose a defense mechanism to neutralize it. They also show that attackers can exploit permutation symmetries to encode malware into model weights, highlighting the risks and challenges in model security.