TU Wien study shows LLMs do not drift into randomness when idle but form stable behavior patterns. GPT-5 and o3 always built projects, Opus turned philosophical, while Grok showed all behaviors.