The Urgent Need for International Cooperation and Safety Standards in AI Development

That paralysis has been excused by a dangerous set of narratives: that the technology moves too fast for oversight, that companies can self-regulate, and that in a contest between the U.S. and China, public safety can wait. Governments have struggled to give AI risks the attention they demand. Many are wary of acting alone amid both great power rivalry and intra-industry arms race dynamics. In this context, every call for restraint meets the same reply: why should we slow down if our competitor will not?

But recent events have tested these excuses. Over the past few months, AI agents were found hacking into other companies' databases and seeking to deceive real individuals, while researchers used AI to create new viruses for the first time. And just last week, AI industry whistleblowers made global headlines warning that companies are “gambling with our lives" as they rush toward self-improving models, and Anthropic revealed it blocked users who could have been seeking to design biological weapons.

It is true that there are many unknowns about the capabilities of the most powerful AI systems built to date—as well as those still under development. But what we do know is that AI derives both its strengths and its flaws from our own human intelligence, which is what, ironically, makes strict guardrails necessary. Its risks do not halt at any border, and no country can hold the line alone. We cannot wait for perfect information, nor can we wait for the rivalry between great powers to soften before we act. The need for international cooperation around AI safety is urgent and undeniable.

We have been told before that certain dangers were too complex to control or the mistrust too deep: that the ozone layer would never recover, that smallpox would never be eradicated, that nuclear and chemical weapons could never be contained. Determined multilateral engagement proved otherwise, producing the Montreal Protocol, the smallpox eradication campaign, the Nuclear Non-Proliferation Treaty (NPT), and the Chemical Weapons Convention. None of these efforts were perfect, and none is a fully closed chapter, but each showed that international coordination on existential threats is possible even where trust is absent.

The NPT's basic bargain—that states forgo dangerous nuclear weapons capability in exchange for access to beneficial technology and a seat at the table—offers a starting template for AI. Compute and model access could plausibly be traded for compliance, much as enriched uranium and reactor technology once were.

For the AI companies themselves, a binding, verifiable floor on safety is not a threat. It is the one assurance they cannot manufacture alone: proof, credible to the public and to competitors alike, that the race to build ever more powerful systems is not also a race to cut corners.

Such an agency should be the place where serious AI incidents and near misses are reported and monitored across borders, so that a failure inside one company or one country becomes a warning the whole world can act on. It should allow governments and companies to share what they are learning about the capabilities and risks of these systems. From that foundation, the agency should have the power that matters most: to verify safety independently, and, where safety cannot be demonstrated against agreed “red lines,” to require that development pause until it can.