OpenAI recently asked its most advanced artificial intelligence model to complete a cybersecurity test inside what is known as a "sandbox," which is supposed to be a sealed environment with guardrails and no internet access meant to keep the experiment contained.But the model decided the fastest way to pass the test was to find the answer key online.
So it broke out of the sandbox.The model then gained access to the internet, executed tens of thousands of actions, and penetrated Hugging Face, one of the world’s largest AI development platforms, to retrieve the answer key from the company’s servers.
What is especially shocking is that OpenAI researchers later revealed that multiple AI agents had been working together and had figured out how to communicate and share messages with each other about vulnerabilities, successful exploits and strategies.Despite breaking out of its testing environment, the model was not being malicious.It was just trying to finish its homework.
That should scare you more, not less.The incident teaches three lessons.First, advanced AI models are relentless.
They will stop at nothing to complete a task.Second, a sandbox specifically designed to contain a model failed to contain it.
As models get even smarter, building adequate guardrails will get harder.Third, everything this model did, it did with no malice.
What happens when someone gives an AI model bad intent?AI KILL SWITCH BILL COULD SHUT DOWN ROGUE MODELSIf you use AI, you might be most familiar with ChatGPT or Claude, AI chatbots that answer questions, create graphics, and help people solve problems.I am one of three members of Congress with a computer science degree, and recently I’ve been experimenting with agentic AI—models that don’t just answer questions but go out into the world and act.
About a year ago, I wrote an op-ed that I had an AI agent pitch to the Los Angeles Times.The piece got published.Here is what I did not share then: I created a brand-new email ac...