A new simulation finds that an AI agent's long-term safety depends less on its core programming and more on the environment and other agents around it.
