Don’t Freeze Progress — Harden Defenses: Why a Blanket AI Pause Would Hurt U.S. Leadership

Senior AI leaders have warned that current safeguards lag behind what advanced systems can do. But instead of pursuing a broad, hard-to-enforce pause in frontier-model development, the United States should double down on concrete defenses: clear red lines, stronger model safeguards, hardened critical infrastructure, defensive AI investments, and large-scale workforce training.

Why the Pause Debate Matters

On September 12, Anthropic CEO Dario Amodei urged a measured slowdown in the pace of capability improvements for the most powerful AI models, warning that coordinated malicious agent swarms could cause internet-scale damage within months. Sam Altman and Elon Musk publicly supported the call. Amodei emphasized he was not asking labs to stop training models, only to decelerate capability gains — but a coordinated pullback by leading labs would effectively slow overall progress.

On September 23, during UN General Assembly week, Amodei and Altman presented their concerns to the UN Security Council, calling for shared testing standards and an international incident-reporting mechanism. The White House science adviser in that forum countered that rapid frontier progress alone is not a reason to pause development or to erect new global governance structures around it.

Why a Blanket, National Slowdown Is Ineffective

AI is fundamentally different from countable, physical technologies like nuclear materials. Model weights and software can be copied, modified, and redistributed. One team’s system can enable both beneficial breakthroughs and malicious misuse. That fungibility makes a national or regional pause a blunt tool: determined bad actors can continue development and deployment elsewhere.

“Inspections and monitoring helped deter nuclear proliferation, but software can be cloned and repurposed.” — Key practical distinction raised at the UN session.

Practical realities also weigh against a long pause. Gartner estimates task-specific AI agents will be embedded in roughly 40% of enterprise applications by the end of 2026. Those agents already access internal data, software, email and operational tools — functions businesses cannot simply freeze for weeks or months without disruptive consequences.

TOPSHOT - China's President Xi Jinping proposes a toast during a State Dinner hosted by US President Donald Trump in the East Room of the White House in Washington, DC, on September 24, 2026. President Xi is on a State visit to the United States from September 23 to 25. (Photo by Brendan SMIALOWSKI / AFP via Getty Images)

What the U.S. Should Do Instead

Rather than an across-the-board halt, U.S. policy should emphasize actionable, enforceable measures that reduce risk while preserving leadership:

• Set Clear Red Lines: Define and publish categories of dangerous capabilities and high-risk deployment contexts that require special controls or prohibition.
• Strengthen Frontier Safeguards: Require rigorous third-party testing, adversarial evaluation, and mandated incident reporting for the most capable systems.
• Harden Critical Infrastructure: Fast-track programs to protect energy, communications, finance and other critical systems from AI-enabled attacks.
• Invest in Defensive AI: Scale public investment in AI tools for cybersecurity, fraud prevention, threat detection, and resilience.
• Build Secure Infrastructure: Expand reliable electricity, secure compute capacity, resilient data centers, and domestic semiconductor supply chains.
• Workforce Development: Launch nationwide upskilling, reskilling and training programs so humans can operate, audit and govern advanced systems.

Roles for Industry

Frontier companies must implement immediate, verifiable internal guardrails. That includes testing models for biological and chemical misuse, simulating adversarial scenarios (including fraud, impersonation, and operational abuse), and developing robust incident-response processes. Companies should adopt the principle that systems with catastrophic-risk profiles are not trained or deployed without adequate safeguards.

Geopolitics and the Global Race

Competitors matter. China’s “AI Plus” agenda emphasizes aggressive, state-led deployment across science, industry and public services. Policymakers cannot assume global cooperation; U.S. strategy must be resilient to rivals who may advance without similar safety commitments. Leadership requires control of the entire stack — compute, semiconductors, energy, data centers, talent, and secure distribution — not just a leading model.

Conclusion

A targeted, accelerated program of safety, defense and innovation is not a retreat. It is the practical path to maintain American leadership while preparing for catastrophic misuse. Pause decisions remain appropriate on a case-by-case basis when a frontier lab cannot demonstrate safety; but a blanket industry-wide slowdown would be difficult to enforce, damaging to industry and potentially harmful to national security. The urgent task is to build multi-layered defenses, clarify regulatory red lines, and ensure both industry and government act now.

Originally published on Forbes.com