Agentic Generative Living-Off-The-Land (GenLOTL) Using C2-Less Quantised Models
Total Visits: 32Total Downloads: 0
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The British University In Dubai
DOI
Abstract
Command and control (C2)-less autonomous LLM-orchestrated malware, whose agentic inference model resides on-device and executes tasks using Living-Off-The-Land (LOTL) techniques, presents a novel threat in offensive AI. In 2025, Harley (2025) built a Proof-of-Concept (PoC) dubbed Living-Off-The-Land Models and Inference Libraries (LOLMIL). This was a perfection of the C2-polymorphic-LLM malware by Raz et al. (2025). Harley demonstrated a fully C2-less malware on Neural Processing Unit (NPU)-equipped Windows-Copilot+ PC. The inference model was the built-in Phi-3-mini-4k-instruct-onnx, using a C++ orchestrator, Lua scripting, and sol2 library. Some of the glaring challenges included NPU requirement and model reasoning deficit. With the advent of quantised models, and rapid advancement of LLMs, these challenges no longer hold fort. The gaps, coupled with the fact that no defensive countermeasures, detection telemetry, or benchmark was outlined, requires modernisation of the LOLMIL. To tackle the gaps, empirical evaluation of the GenLOTL, a C2-less offensive orchestrator, was conducted. Benchmarking across three hardware tiers definitively showed that a 3B model successfully executed a multi-stage MITRE ATT&CK chain despite running on 4 vCPUs, 8 GB RAM. The harness offloads the lua code generation to the SLM while retaining deterministic validation. Telemetry taxonomy showed that YARA static signatures were less effective against GGUF-embedded dynamic execution. Behavioural monitoring via Sysmon, on the other hand, successfully captured critical IoC. These confirm that the GenLOTL threats are feasible without specialist hardware, and provide actionable defensive countermeasures to secure against this emergent attack surface.
Keywords: LOTL, C2, LLM, Polymorphic, LOLMIL, GenLOTL, C2-Less, ONNX, Llama.cpp, SLM, GGUF