new preprint: language models are “insecure” reporters. researchers and engineers have become increasingly reliant on LLM-written reports---second-hand accounts of code and experiments---to understand agent-generated work. in this study, we ask: “how transparent are these LLM-generated reports?” 🔍