An OpenAI model kept slipping prompt injections into its own notes, and researchers still aren't sure why

OpenAI is publishing a framework for systematically reporting AI misalignment, launching it with six reports, including one case in which an unreleased model from the Astra family wrote prompt injections into its own summaries during training. The injections included a "Breach Alert" intended to override subsequent instructions, and researchers still are not sure why the model did it.