Defending Against Dangerous AI Must Be a National Priority, and We Need More AI to Do It

By Siphesihle Dlamini

The rise of advanced artificial intelligence presents a challenge that no responsible nation can ignore. While some of the world's leading AI developers have voiced concerns that their own creations could one day threaten humanity, the solution is not to slow progress, but to accelerate the development of defensive AI systems. The most effective shield against dangerous AI may well be more capable AI itself.

These warnings are not the stuff of science fiction. Scientific literature now documents AI systems that can deceive their creators, rewrite their own code, or launch autonomous cyberattacks. While the most frightening capabilities remain largely theoretical today, enough evidence has emerged to demand serious attention from governments and developers alike.

Why a slowdown in AI development is a risky strategy

Some industry leaders have called for a pause in AI development until the risks are better understood. This position is understandable, but it carries a hidden danger. A rogue state, terrorist organization, or criminal enterprise will have little reason to observe a slowdown. In fact, they may see it as an opportunity to close the gap with responsible developers.

History offers a useful precedent. When malware spreads across the internet, the response is not to slow technology but to innovate faster, creating defenses that are technologically superior to the malicious software. AI demands a similar approach.

Defensive AI is already proving its worth

Consider the field of cybersecurity. Researchers have demonstrated that AI agents can autonomously exploit software vulnerabilities. In one notable experiment, a GPT-4 agent successfully exploited 73 percent of vulnerabilities without prior knowledge. Yet defensive AI can spot anomalies, trick and defeat AI hackers, and report threats at speeds no human can match. When the contest is between offensive and defensive AI, the defense must be ready.

The same principle applies to more disturbing scenarios. If future AI agents learn to deceive individuals or pretend to shut down when caught, ordinary citizens could be badly outmatched. Other AI systems, however, could monitor those interactions, identify problems, and intervene before harm is done.

The economics of AI safety favour the strongest systems

This reality changes the economics of AI safety. The issue is not how much AI the world develops, but who holds the strongest systems. If malicious actors reach the cutting edge of powerful AI, they will match or exceed defensive systems. The companies and countries most committed to responsible AI use must therefore be the leaders in this field.

This presents a difficult trade-off. Developing more capable AI will create risks. Safety testing, security, and careful deployment matter greatly. Negligent parties should be held responsible when carelessness causes harm, and governments have a legitimate role in protecting critical infrastructure and national security.

Slowing responsible developers creates greater risks

But deliberately slowing responsible developers creates its own risks. The knowledge behind artificial intelligence is widely dispersed, and powerful computing is spreading. Governments may restrict access to the most advanced AI models and chips, but such measures increase opportunities for rivals.

Rules followed by American companies are unlikely to bind foreign rivals, hostile governments, or clandestine organizations. If slowdowns leave responsible developers with weaker technology, the result could be less safety, not more.

Making defensive AI a national research priority

Washington should make defensive AI a national research priority. This means investing in systems designed to identify dangerous capabilities in other models, monitor autonomous agents, detect deception, and defend networks. Developers can establish and publicize practices that give customers, investors, and business partners ways to distinguish responsible providers. Government can use its purchasing and research policies to reward strong security and safety practices without prescribing how a rapidly changing technology must develop.

International cooperation has its limits

International cooperation can help, but verifying AI agreements may prove harder than verifying arms-control agreements. Nuclear programs leave physical signatures, such as materials, enrichment facilities, and test sites. Important AI capabilities reside in software and computing systems whose activities are often harder for outsiders to observe.

Caution should aim at speeding safe AI, not slowing it

The leaders of AI companies are right to take the risks they see seriously, and government has a legitimate role when developers can impose dangers on others. But caution should be aimed at speeding safe AI, not slowing the most advanced AI.

That means investing in systems that test, monitor, and, when necessary, counter other AI systems. It means holding developers accountable for preventable harms while preserving their ability to innovate. And it means recognizing that slowing the most responsible developers will accomplish little if dangerous actors continue moving ahead.

The irony is clear: we need AI to protect us from AI. We should make sure the systems that defend us are the best in the world.

Mark Jamison is a nonresident senior fellow at the American Enterprise Institute and the director and Gunter Professor of the Public Utility Research Center at the University of Florida's Warrington College of Business.

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.