Study finds cheap AI-generated malware variants share a detectable signature

An arXiv preprint reports that Claude Opus 4.6 generated structurally diverse data-exfiltration payloads for $0.41-$0.73 per validated payload, but a single static signature detected every generation in the study. Researchers built a pipeline that generated, tested and refined code for file traversal, encryption and exfiltration. Giving the model a structured history of earlier variants produced code that some machine-learning classifiers struggled to cluster. However, unchanged boundaries between modules allowed the common detection signature. A complementary evaluation of seven local small language models, with 20M-1.5B parameters, found they lacked the generative stability needed for reliable production of such variants.