Papers for
cybersecurity defense teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Language model defense makes attackers use more computing energy
Make the Adversary Spend More Energy: Generative Deception against Language-Model Attackers in Critical Infrastructure
Abstract: Language-model attackers consume computational resources as they explore and interact with a target, creating an opportunity for cyber deception to impose costs. Whether such engagements favor the defender requires measuring both sides. We evaluate an autonomous language-model attacker against a decoy that uses a defending language model to analyze attacker commands and select defensive responses. Across 330 sessions, eight model pairs, two attacker profiles and three counter-injection levels, both sides run on identical GPUs and their energy is measured above idle. With runtime thinking disabled, the attacker-to-defender energy ratio has a 95% interval above one in 41 of 48 configurations under an allocation excluding attacker diagnostic calls, and in all 48 when those calls are included. Token ratios remain within approximately a factor of two in configurations without generation-cap truncation, showing that token counts can obscure substantial energy asymmetry. Counter-injection increases attacker tokens by approximately 9% per level without a detectable pooled change in energy consumption. Enabling thinking on both sides increases defender energy by 8--17 times and reverses the exchange for one model pair in the original campaign. We use these findings to motivate a proposed SCADA deception architecture with private physical state per identified adversary and deterministic protocol and process emulation.
Small language models show promise for generating malware code tasks
From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation
Abstract: Small 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. Yet the ability of ultra-compact models to support malware during post-exploitation remains poorly understood. We systematically evaluate seven open-source SLMs (20M-1.5B parameters) on two post-exploitation code-generation tasks: Directory Enumeration and Backdoor Generation. We compare Python (high-resource) and Lua (low-resource) across prompting strategies, task formulations, iterative reasoning, grammar-constrained decoding, and execution-feedback loops. Results show that offensive code-generation capabilities do not increase linearly with model size. Successful execution depends on the interaction of model capacity, task formulation, and prompting strategy, with task formulation exerting a greater influence than prompt engineering alone. Algorithmic formulation combined with few-shot prompting enables Qwen2.5-Coder 1.5B to reach Pass@5 of 100% in both Python and Lua for Directory Enumeration, 100% in Python for Backdoor Generation, and 99.2% in Lua. Although the evaluated models remain too computationally demanding for widespread deployment in contemporary malware, rapid progress in compact models may reduce this barrier. These findings help anticipate AI-assisted malware and inform future countermeasures.