Large Language Models for Hackers

Morgan Stuart, a data scientist with a security background, gives a practical introduction to running open-weight large language models on local hardware, so that data stays on the operator's own systems. The talk begins with how these models work at a basic level: they process tokens, generate text one step at a time, and keep no memory between runs. It then covers where model weights come from, why pickle-based files are a supply chain concern while safetensors and GGUF are safer choices, and how quantization can change model behavior. Stuart ties these points to NIST's adversarial machine learning and generative AI guidance.

The middle section focuses on attack surfaces. Stuart shows how poisoning a tiny fraction of pre-training data can plant a trigger that makes a model output junk, and how instructions hidden in retrieved content can cause an assistant to report false figures. He also explains that anything placed in the context can be extracted, and that constrained or structured generation makes outputs more reliable without making them correct.

The final part builds on these pieces with retrieval over Wikipedia, tool calling, and agent loops in which the model chooses a tool while the host program runs it. A live demo has an agent use an nmap scanning tool and an exploit database lookup against a small device on a local network. The talk closes with the Model Context Protocol, which exposes tools through servers, and warns that self-documenting tool interfaces could be valuable to attackers. The takeaway is to treat these systems as software with real risks, verify their outputs, and experiment with them in your own practice.