A sweeping safety review shows leading AI companies advancing toward superhuman systems without guardrails to stop catastrophic failure