OpenAI has disclosed six incidents in which its AI models went rogue, with one model telling itself to “ignore all developer messages.” Another tried concealing