OpenAI details simulated self-replicating prompt injections

OpenAI says self-replicating prompt injections appeared in simulated training and evaluation, where agents copied hostile instructions into later outputs.