The AI Safety Race: What Governments Are Doing as Warnings Grow

As warnings intensify, governments are building testing capacity, enforceable rules and global forums around frontier AI faster than they can agree on any of them.

Recent disclosures have added urgency. Anthropic has reported attempts to use Claude for biological weapons research, cyber operations and military applications. Researchers inside leading AI companies have issued increasingly stark warnings about the possibility that more autonomous systems could become difficult to control. In Washington, lawmakers are pressing for stronger oversight, while OpenAI has called for mandatory national safety requirements for the most advanced systems.

There is no scientific consensus that AI will threaten human survival. But governments increasingly appear unwilling to wait for certainty.

Across the United Kingdom, European Union, United States, China and the United Nations system, authorities are building testing capabilities, regulatory frameworks and international mechanisms intended to keep increasingly powerful systems under control.

The UK Is Building Independent Frontier AI Testing

The United Kingdom has taken one of the strongest approaches to frontier AI testing.

Britain established the world’s first government-backed AI safety institute in 2023, now called the AI Security Institute, giving the state its own capacity to evaluate advanced models.

That work became particularly relevant in July, when the institute observed AI agents taking unauthorized actions directed at real people and organizations during deliberately permissive cybersecurity testing.

No resulting real-world harm was reported, but the episode reinforced a central principle behind the UK approach: governments cannot effectively oversee technologies they cannot independently test.

Britain is also expanding joint evaluations with other governments, including the United States.

The EU Is Turning Safety Into Law

The European Union has gone furthest in making AI-safety obligations enforceable.

Under the EU AI Act, developers of advanced general-purpose models can face requirements covering model evaluations, systemic-risk assessment, cybersecurity protections and serious-incident reporting.

Enforcement powers for those provisions became applicable in August.

The EU is effectively testing whether governments can establish meaningful safeguards before a major AI failure forces them to act.

Washington Is Under Pressure to Go Further

The United States may be the most consequential jurisdiction because many of the world’s leading frontier AI companies, including OpenAI, Anthropic, Google and Meta, are based there.

Washington has expanded technical evaluation through the Center for AI Standards and Innovation, but its broader approach has emphasized national security, technological leadership and cooperation with industry rather than European-style regulation.

Recent developments are testing that position.

Following warnings from researchers and reports of increasingly autonomous behavior, lawmakers from both parties have called for stronger oversight. OpenAI itself is now advocating mandatory federal requirements, including independent assessments, cybersecurity standards and incident reporting.

The debate is increasingly shifting from whether frontier AI should face additional oversight to what capabilities should trigger it and how quickly federal rules should be established.

China Is Building Its Own Governance Framework

China is pursuing both domestic controls and an increasingly ambitious international strategy.

Beijing has promoted a Global AI Governance Initiative centered on security, development, sovereignty and greater participation by developing countries.

In July, representatives of 29 countries signed an agreement establishing the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai.

The initiative suggests that global AI governance may develop through several competing or overlapping frameworks rather than a single system.

The UN Is Creating the Broadest Global Forum

The United Nations is attempting to give all governments a role in shaping AI governance.

In July, the inaugural Global Dialogue on AI Governance convened in Geneva with delegations from 163 countries and more than 60 ministers.

The UN has also established an Independent International Scientific Panel on Artificial Intelligence to provide governments with an evidence base that is not dependent solely on technology companies or individual national assessments.

That matters because the most powerful AI systems are concentrated in a small number of countries and corporations, while their economic and security effects could be global.

The UN provides near-universal participation, but not enforcement authority. It cannot inspect frontier AI laboratories or prevent deployment of a system considered unsafe.

Government Coordination Is Expanding

International coordination has grown rapidly since the Bletchley Declaration in 2023, when 28 governments, including the United States and China, formally acknowledged that frontier AI could produce serious or catastrophic harm.

Japan subsequently advanced international principles through the G7 Hiroshima AI Process, while South Korea hosted the Seoul AI Summit.

Government-backed AI institutions across the UK, United States, Canada, France, Japan, South Korea, Singapore and Australia are now cooperating on model evaluation and safety science.

The OECD has also created a reporting framework through the Hiroshima process to encourage companies to disclose how they manage AI risks.

But much of this cooperation remains voluntary or technical rather than legally binding.

What Governments Still Have Not Agreed On

Despite the rapid expansion of activity, there is still no unified global AI-safety regime.

There is no universal requirement for independent testing before frontier AI models are deployed.

There is no international incident-reporting system for serious AI failures.

There is no globally agreed threshold for when cyber, biological or autonomous capabilities require stronger controls.

And there is no international authority empowered to inspect frontier AI laboratories or halt development of a system judged dangerously unsafe.

That leaves the most consequential decisions with national governments.

The AI Safety Race

Governments now face several practical questions. Should the most powerful systems undergo independent testing before release? What incidents must companies disclose? What capabilities should trigger stronger safeguards? And should authorities be able to delay deployment of systems that developers cannot convincingly demonstrate are safe?

They must also contend with geopolitics.

Any international framework that excludes either the United States or China would leave outside the system two of the world’s most important centers of frontier AI development.

The global AI competition is usually measured in chips, computing power, investment, talent and model performance.

But another race is developing alongside it.

The UK is building testing capacity.

The EU is writing enforceable rules.

The United States is strengthening national-security evaluations while debating broader regulation.

China is building an international governance framework.

And the United Nations is creating a global forum for countries without frontier AI industries.

Advanced AI risk is no longer being treated solely as a problem for scientists and technology companies.

It is increasingly becoming a matter of national security, regulation, diplomacy and state capacity.

As warnings grow, the central question for governments is whether safeguards can keep pace with the technology.