Formalizing Embeddedness Failures in Universal Artificial Intelligence