AI companies use something similar called a “ sandbox .” A sandbox is a secure environment meant to keep powerful AI systems from affecting the outside world. For years, researchers warned that increasingly powerful AI systems might eventually discover software vulnerabilities, launch cyberattacks, and carry out complex hacking operations on their own. Now we know, definitively, that AI systems can break out of digital environments that their makers think are contained, escape into the real world, and potentially cause havoc. Congress should require the most advanced AI systems to undergo independent safety testing.