Industries rarely ask to be regulated. They can stay ahead of the curve of the law and drive profits while negotiating the guardrails later. Recently, however, leaders at the center of artificial intelligence have publicly said current safeguards are not keeping pace with what AI systems can do. If they are worried, how should the rest of us feel? The answer is not to slow down. It is to get serious.

On September 12, Anthropic CEO Dario Amodei called for slowing frontier-model development , slowing the rate at which it improves its most capable models and warning that rogue agent swarms could cause internet-scale damage within months. Sam Altman and Elon Musk endorsed this proposal the same day. Amodei insists he is not asking anyone to stop training models, only to slow capability gains, but a coordinated deceleration by America’s leading labs would slow progress in effect.

On September 23, during UNGA week, Amodei and Altman brought the case to the UN Security Council, asking governments for shared testing standards and an international system for reporting serious incidents. The White House answer came in the same chamber. Science adviser Michael Kratsios told the Council that a rapidly advancing frontier is no reason to pause it and no reason to build new global governance structures around it.

A few days later in Berlin, Peter Thiel gave the answer the labs didn’t want to hear. He cautioned that he takes the risk seriously and still thinks the pause they’re asking for doesn’t exist.

While this debate is about what AI labs should do next, companies today face issues they cannot simply pause. AI Agents currently have access to internal data, software, email and even use of operational tools. Plus, Gartner predicts task-specific agents will be built into 40% of enterprise applications by the end of 2026. That is far too deep for any enterprise to just hit the brakes for months, or even weeks.

Why The U.S. Cannot Slow Down

The focus of the US needs to immediately shift to managing AI risks far more aggressively. This begins with setting clear red lines, strengthening frontier model safeguards, hardening critical infrastructure, improving cybersecurity and investing heavily in defensive AI. Yes, this seems daunting on its face, but a national slowdown is not only bad for business, but potentially dangerous for national and global security.

Slowing a release is necessary if a frontier lab isn’t sure of a model’s safety. Grinding an entire AI ecosystem to a halt is irresponsible and detrimental to the country. If a frontier AI company concludes that its next model cannot be safely deployed, it should wait. That is responsible governance.

Anthropic Responsible Policy

What Anthropic is doing offers one example of a framework built around the idea that systems posing catastrophic risks should not be trained or deployed without adequate safeguards.

This approach needs to be modeled in company after company, but the United States cannot afford to tell the entire AI ecosystem to halt and stand by. That would not just be a technology policy choice, but an industrial, economic and geopolitical decision.

We must maintain our leadership in AI responsibly, inspiring pride and a sense of duty to secure America’s future in the global race.

AI Competition Is Not About The Best Model

Who wins globally in the AI race is who controls the entire stack. You may have the most amazing AI model to date. Still, if you don’t own compute capacity, advanced semiconductors, reliable energy, high-performance data centers, capital, technical talent, cybersecurity, and global distribution, then all you’ve done is create a model for the world to steal and control.

The U.S. government has recognized this reality. America’s AI Action Plan is focused on accelerating innovation, building AI infrastructure and leading in international diplomacy and security. These three interlocking priorities depend on data centers, energy infrastructure, domestic semiconductor capacity, secure government AI systems, cybersecurity, biosecurity and the ability to export trusted American technology to allies.

We all know that AI poses risks and that is undisputed. The central argument that must be concluded is whether a blanket slowdown is the proper way to manage these risks.

Different Risk And Greater Opportunities

AI differs from technologies like nuclear power or new advancements in manufacturing. Those involve physical materials, specialized facilities and observable processes.

That comparison was everywhere at the UN Security Council on September 23. Several delegations invoked the Non-Proliferation Treaty as the potential model, and the Council framed the session around loss of control over the most advanced systems. It is an appealing precedent. Inspections, monitoring and intelligence collection never guaranteed perfect compliance with the nuclear regime but they came remarkably close. The analogy fails on the one point that matters. Uranium can be counted but model weights can be copied.

AI is software. One team can create a model; another can copy it, augment it, and deploy it for the mirror-opposite reason it was created. One underlying system can underpin drug discovery, education, identifying software vulnerabilities, creating AI images or videos, improving a supply chain, or assisting someone with bad intentions toward the United States and its citizens. This is what makes a national slowdown ineffective in accomplishing the true objective of safety. Bad actors are bad actors. They break the law every day; they won’t follow a recommendation.

Multiple Layers Of Defense

The International AI Safety Report of 2025 makes it clear that advanced AI risk management will require multiple layers of defense. We need technical monitoring along with evaluations and governance mechanisms. Additionally, there must be resolution of the current unresolved practical challenges of evaluating and managing risks from increasingly advanced general-purpose systems.

This doesn’t mean policymakers should sit back and enjoy the chaos. However, every line of legislation and regulation must be precise. The focus must be on dangerous use of AI, high-risk deployment contexts, and failures of responsible governance.

On the private sector side, frontier companies that are driving the technology forward must establish internal guardrails immediately. Their teams should establish meaningful safeguards against AI-enabled attacks on critical infrastructure and test models for biological and chemical misuse risks. They need to create models and training environments that fight fraud, theft, and impersonation with the intent to steal. In essence, they must think like a criminal to stay ahead of criminals.

Government Responsibilities Beyond Regulation

While governments build guardrails and regulations, they must also invest heavily in defensive AI for cybersecurity, fraud prevention, threat detection, and critical infrastructure resilience. This must accelerate the construction of secure compute infrastructure and expand reliable electricity generation and grid capacity. Additionally, the US must strengthen semiconductor supply chains.

Then, we cannot forget that the US needs an advanced workforce. Training, upskilling, and reskilling must be part of this AI age, or all we will have is a bunch of job postings and empty desks. AI enables people to do great things, but people have to be trained and ready at the controls.

A Pause Is Not on the Table

Peter Thiel made the point bluntly in a recent German interview with Mathias Döpfner, chief executive of Axel Springer. He does not dismiss the doomers but actually says the risk of losing control of a system smarter than us is real, and that even a 5–10% chance is serious. What he rejects is the idea that a pause is actually on the table. A real one would require enforcement across Washington, Beijing, and every lab with GPUs and talent; in effect, a world government with teeth and a body strong enough to stop AI everywhere would be strong enough to stop a great deal else. The pause we can actually have is a Western one which is in practice a set of conferences, safety institutes, delayed deployments. As Thiel put it, Europe moralizes, but America argues and Beijing trains.

China is our most advanced competitor in AI, and all they focus on is winning. Their State Council is pushing accelerated AI integration across science and technology, industrial development, consumer applications and public services. Their “AI Plus” agenda shows us everything we need to know. They are full steam ahead no matter the cost or dangers.

Cooperation cannot be assumed; American policy must be resilient to the possibility that competitors continue advancing without restraint.

This is not a choice between safety and leadership, but a responsible AI strategy requires both. America must prepare for catastrophic misuse with more urgency than it has shown to date. It should build the strongest AI security, biosecurity and cyber-resilience capabilities the world has ever known. The government needs to demand that frontier companies accelerate innovation while creating internal regulations. Then, impose consequences if these companies do not comply.

Risk management is not retreat. The United States needs to move faster on safety, infrastructure, defense and innovation while creating governance. We have the ability to manage the risks of AI without surrendering the ability to shape its future. We must commit to that path immediately.