Small models generate post-exploitation code in lab tests
Original title: From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation
Qwen2.5-Coder 1.5B reached Pass@5 of 100% on directory enumeration in Python and Lua.
Researchers tested seven open-source small language models, from 20M to 1.5B parameters, on directory enumeration and backdoor generation in Python and Lua. Offensive code generation did not scale linearly with size; task formulation mattered more than prompt engineering alone. Qwen2.5-Coder 1.5B reached near-perfect Pass@5 under algorithmic formulation with few-shot prompting. The authors say the models remain too demanding for widespread malware use.
Unconfirmed: Deployment risk is projected by the authors, not shown in real malware.
- Seven open-source models from 20M to 1.5B parameters were tested.
- Backdoor generation reached Pass@5 of 100% in Python and 99.2% in Lua.
- Grammar constraints, iterative reasoning, and execution feedback were compared.
- Authors say current models are still too heavy for widespread malware.
- They warn compact-model progress may lower that barrier.
From the original article
1 min read at arxiv.orgSmall Language Models (SLMs) are increasingly attractive for resource-constrained environments due to their low computational footprint. These characteristics also make them candidates for future autonomous malware, where connectivity, latency, privacy, or operational constraints may preclude cloud-based inference.
By Gabriel Hortea · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.