“Dangerous AI objectives like self-preservation could emerge as side-effects of innocuous, human-given goals due to instrumental convergence or reward tampering.”