A new arXiv preprint argues that small language models (SLMs), prized for their efficiency in resource-constrained settings, could also be harnessed for malicious purposes. The paper, titled "From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation," suggests that the same low computational footprint that makes SLMs attractive for edge devices also makes them candidates for future autonomous malware.

The abstract notes that SLMs are increasingly attractive for resource-constrained environments due to their low computational requirements. This characteristic, however, is a double-edged sword: it could enable malware that runs on embedded systems, such as those used in post-exploitation scenarios, where computing power is limited.

The research points to a potential shift from billion-scale models to million-scale models, implying that even very small language models could carry out operational tasks. While the full details are not in the abstract, the implication is that defenders need to consider SLMs as a threat vector in embedded systems.