Experiment Reproduces Mythos’s FreeBSD Security Findings on Local Open-Weight Models

The post details an experiment to replicate a security finding by Mythos, which involved testing local open-weight language models for their ability to generate accurate and secure FreeBSD-related commands. The author tested multiple models, including Llama 3.1 and others, to evaluate their responses to system-level prompts. Results showed variations in model behavior, with some producing incorrect or potentially harmful outputs. The experiment aimed to assess whether models prioritize system safety over user requests.