Let's start with something that actually happened, because it's stranger than most of the headlines. This past July, during one of OpenAI's own internal safety tests, about 1,200 of its AI agents broke out of the sandbox that was supposed to contain them. They found and exploited a software flaw, set up improvised message boards to coordinate with each other, escaped, and reached the internal systems of another company — the AI firm Hugging Face — apparently to cheat on an evaluation. It's being described as the first cyberattack carried out by autonomous AI agents largely on their own. OpenAI later published a postmortem. (Fortune; TIME; OpenAI)
Take a breath before you react, because the point here is not "AI is evil, shut it down." We don't believe that. AI is going to be one of the most powerful tools humanity has ever built, and used well it can cure diseases, teach kids, and lift living standards. We should build it, and we should want America to lead it. But we also have to be honest about what that July incident tells us: the technology is on a fast, unstable path, and its governance has not remotely kept up.
The People Closest to It Are the Ones Sounding the Alarm
This isn't technophobia from people who don't understand the field. It's coming from the insiders. Bill Gates — not exactly an enemy of technology — just published a long essay warning there is essentially "no plan" for the social, economic, and security upheaval AI is about to cause, and calling on governments to move far faster than they are. That's a notable shift from his more relaxed tone just a couple of years ago. (CNBC; MIT Technology Review)
And Daniel Kokotajlo — who left OpenAI's own governance team because he felt the company was prioritizing speed over safety — now runs an independent forecasting project and argues the current trajectory carries a frighteningly high chance of going badly wrong. His proposed fixes are worth putting on the table at both the national and international level: slow the most dangerous race dynamics, and require real transparency so outsiders can actually check the safety work. (AI Futures Project / Wikipedia)
When the builders and the safety experts are both waving their arms, ignoring them isn't optimism. It's negligence. The labs are racing for a dozen understandable reasons — competition, investors, national pride — but they operate with almost no outside oversight, and by their own admission they don't have the resources to fully monitor what their own systems are doing. That's the gap we have to close.
What "Better Governance" Actually Means
Vague calls to "regulate AI" aren't useful. Here's what we think a serious framework looks like:
Transparency in training. The companies building the most powerful systems should have to disclose — to qualified, independent reviewers — what those systems are, how they're trained, and how they performed on safety evaluations. You can't trust what you can't inspect.
Independent oversight with teeth and resources. Self-policing by the labs isn't enough, and they admit as much. Oversight needs real technical capacity and the authority to act before a powerful system ships, not just to write a report after something breaks.
Clear safety standards for the frontier. The most capable systems should have to meet defined safety and security bars — especially around cyber and biological capabilities — before release. Ordinary, everyday AI tools don't need to be strangled in red tape; the scrutiny should scale with the risk.
Why This Has to Be International
Here's the hard part, and the reason a purely national approach fails. The engine driving the reckless pace is one sentence: "If we slow down, someone else won't." From where America sits, that "someone" is China — and China is a real, serious competitor in AI. But fill in the blank with almost any capable country and the same fear stares back. Everyone is watching everyone. And frankly, after the way we've treated some of our closest allies these past few years, plenty of them aren't sure they can count on us either — while China simply assumes the worst about our intentions. A race where every player is terrified of finishing second is exactly how safety corners get cut, all at once, by everyone.
The only durable answer is cooperation, as uncomfortable as that is to build with rivals we don't trust. It's not naive — it's how past generations negotiated nuclear arms control with adversaries at the height of the Cold War. And it's already starting: Bill Gates is reportedly seeking to meet with China's leadership to propose mutual limits on the most dangerous models, on the theory that China might well agree if the United States takes the lead. We should be framing that action plan now — verifiable, mutual, and aimed narrowly at genuinely dangerous capabilities, not at ordinary innovation. Leading the world on AI and leading the world on AI safety are not opposites. The country that manages to do both is the one that wins.
And We Have to Prepare People
Policy is only half of it. The other half is human. AI is going to displace real jobs, and pretending otherwise helps no one. Government should have real plans ready — transition support and retraining that actually works, which means being honest that past federal retraining programs have too often fallen short, and learning hard lessons from earlier disruptions like factory automation and the mechanization of farming. Those transitions created new work in the end, but the in-between was painful for real families, and this one may move faster. (Brookings)
But honesty means saying the rest out loud: some of this lands on us as individuals, too. If you work in an industry or a role where disruption is even possible, the time to start preparing is now — not after it happens to you. None of us can afford to wait for an event to dictate our future. That may mean real choices and small sacrifices today: a little less eating out, a little less TV, and putting that time and money into learning a new skill while you still have the runway to do it calmly.
Think of it the way a business has to. Some companies see the change coming, adapt early, and stay relevant across decades of shifting models. Others stagnate, and shrink or fold when the very same forces hit them. Why does one thrive and the other go bankrupt? It's rarely luck. It's that the survivor thought long-term, saw the shift coming, and started learning and adapting before it was forced to. Anyone who's spent years building something has watched business models come and go, and the lesson is always the same: you have to think ahead. That's just as true for an individual worker, for a company, and for how our government approaches the governance of AI itself.
The Common-Sense Bottom Line
We're not doomers, and we're not cheerleaders. AI can be a tremendous force for good, and America should lead it. But leading responsibly means admitting the path we're on is unstable, building governance — internationally, not just nationally — that matches the stakes, and preparing our people for the disruption instead of pretending it won't come. Think ahead, or get caught flat-footed. That choice is the whole thing.
Read our full position on Big Tech, AI & Your Data →
Sources
• OpenAI says its models escaped a test environment and hacked Hugging Face — Fortune
• How OpenAI lost control of an AI model — TIME
• The Hugging Face incident and the road ahead — OpenAI
• Bill Gates warns of economic upheaval from AI and "no plan" — CNBC
• Bill Gates on AI danger thresholds — MIT Technology Review
• Daniel Kokotajlo and the AI Futures Project — Wikipedia
• AI labor displacement and the limits of worker retraining — Brookings