LLMs worse at proactive investigation than random heuristics