Security researchers have spotted ransomware that uses a local LLM to create variable output from prompts, making it harder to detect.