modified qwen 3.8 27b modifies windows credential dumper to bypass EDR detection

link to original article

TLDR: security researcher eddie zhang used a modified+uncensored local qwen 3.8 27b to create an executable capable of dumping LSASS memory for credential harvesting while evading 2 modern EDR security products.

this makes me reflect on how cloud providers keep putting guardrails on everything to the point where even authorized testing gets blocked. local models are the only real option if we want total control, but running heavy local rigs for long agent tasks drains so much compute and management overhead.

been using claude code hooked to sumus to handle my local project workflows and orchestrate tasks in the background, while keeping full local file access on my machine. curious if anyone here is running local uncensored models as local agents for heavy automation, or if you still blend cloud models with local execution setups for your dev environment?