A new watchdog is tracking when AI goes rogue, and more than 300 incidents were reported in July

AI models lying, scheming, and bypassing safeguards nearly doubled in July, according to new research, including a real hacking campaign by Anthropic and OpenAI models.