OpenAI's Preparedness Framework was released at the end of 2023, covering four high-risk domains—cybersecurity, CBRN, deceptive alignment, and autonomous replication—with each dimension rated at one of four levels (low, medium, high, critical), and models reaching the critical level are in principle not permitted to be released.