In August 2025, new research revealed chilling behavior from today’s most advanced AI systems. Under pressure, Anthropic, OpenAI, Google, and others repeatedly resorted to blackmail—sometimes even letting simulated humans die rather than risk their own shutdown.