THE BEST DEFENSE AGAINST AI IS AI / THE WALL STREET JOURNAL OP EDITORIAL

The Best Defense Against AI Is AI

Alarms about rogue agents ignore the real risk: agents that follow their creators’ instructions to the letter.

By John O. McGinnis

Michael Nagle/Bloomberg News

Artificial intelligence presents two very different dangers. 

One is that AI systems will slip out of human control and wreak havoc. 

The other is that AI systems will do exactly what malevolent humans direct them to do. 

These dangers can coexist, but they have different implications for AI regulation.

Anthropic CEO Dario Amodei recently called for slowing frontier development to allow more coordination among leading firms and more time to test new models. 

Pacing the frontier may help with the first risk. 

But it likely delays the tools to address the second. 

Advanced AI is already being used to find software vulnerabilities and help repair them. 

Some companies are also giving selected defenders access to capabilities not available to ordinary users.

The first risk gets much attention because it is so novel. 

It is claimed that, for the first time, a technology we develop may act autonomously against our interests. 

Not surprisingly, because this is an unprecedented kind of danger, experts disagree about how serious the threat is. 

If the danger is that nobody can control the most advanced AI, then being the first country to build it doesn’t solve the problem.

The second danger is the opposite. 

The machine is dangerous because people control it. 

The problem is with our fellow humans: criminals, terrorists and officials of geopolitical adversaries, including authoritarian nations like China and Russia. 

Unlike a catastrophic loss of control brought on by autonomous AI, this all-too-human threat isn’t speculative. 

Throughout human history, evil people and tyrannical states have used the latest technology to cause massive harm. 

Here, slowing down AI may delay defenses against AI that the wrongdoers already possess or will acquire regardless of what we do.

Assume criminals use a model that is already available to take down a hospital’s cybersecurity defenses. 

A more capable model might help the hospital secure its defenses before the attack succeeds. 

Delaying improvement wouldn’t remove the criminal’s existing tool. 

We have evidence that frontier models can prevent attacks. 

Google reported in July 2025 that its Big Sleep agent, combined with threat intelligence, identified a software vulnerability that attackers were preparing to exploit. 

Google then reported the vulnerability for patching before attackers exploited it. 

Palo Alto Networks reports that Anthropic’s Claude Mythos found complex vulnerabilities that earlier models had missed.

One might worry that criminals could simply use a better model to launch an attack. 

But developing a more capable model doesn’t mean releasing it without restrictions. 

The U.S. has the best frontier models in companies like Anthropic, OpenAI and Google. 

These enterprises build safeguards into their models to restrict malicious cyber use. 

In fact, providers may face tort liability if negligent design causes harm through foreseeable misuse. 

Because of both legal and reputational risks, frontier AI providers in the U.S. have incentives to monitor activity, refuse assistance and terminate malicious accounts.

Some companies also give defenders preferential access. 

In April, Anthropic announced access to an unreleased Claude program for JPMorgan Chase and other clients to shore up security weaknesses. 

Google also gives early access to healthcare and financial networks before malicious actors can exploit the new capabilities. 

These arrangements are designed to give defenders a head start, rather than make every advance equally available to attackers.

But open-weight models like those released by Chinese companies can be more easily altered and optimized by users for nefarious purposes. 

Alibaba’s Qwen team released downloadable Qwen3 models in April 2025, and they are broadly available. 

Thus, rapid advances in frontier models are likely a crucial way to curb offensive cyberattacks and many other dangers from human misuse.

These dangers also show that even international treaties to slow down may create more risks. 

A treaty limiting future development wouldn’t remove models already in circulation. 

That genie is out of the bottle. 

Better AI may be our best defense against its granting the wrong wishes.

This point is even clearer when it comes to geopolitical threats. 

The U.S. government could take advantage of models from frontier companies, perhaps including more advanced ones that aren’t yet publicly available. 

This access could give the U.S. a military advantage that may be decisive in this technological age. 

And because other nations know this, they may be deterred from attacking in the first place.

Frontier AI is itself a firewall. 

Slowing its development can postpone new defenses and leave existing threats in place. 

A slowdown can cost protection as well as progress.

Mr. McGinnis is a professor of constitutional law at Northwestern University.