In our recent discussion of the federal debt, we considered whether stronger productivity growth could help the economy grow faster. Artificial intelligence is one possible source of that growth. That raises a broader question of its own: how much could AI increase productivity and opportunity, and what risks come with it?

The opportunity is enormous.

Productivity means producing more output from the same labor, time and capital. A small company can already use AI to research markets, analyze documents, write software, answer customers, translate material, design advertising and handle administrative work. Tasks that once required programmers, analysts, researchers, designers, marketing staff, administrators and outside consultants can increasingly be done by a handful of people with the right tools.

But AI could also go wrong. It can be used for fraud and cyberattack. Highly autonomous systems can behave in ways their creators never intended. Governments can use it for surveillance. And regulations written in the name of safety could protect today's largest AI companies from tomorrow's competitors.

So the question is not whether AI is good or bad. It is this:

Who will control this extraordinary new productive power—and how do we guard against genuine dangers without suppressing competition and innovation?

AI can make the little guy bigger

One of AI's most important economic effects may be to reduce the advantages of size.

A five-person company cannot afford a large staff of specialists. Increasingly, it can buy pieces of those capabilities through AI. That lowers the cost of entering a market and lets an entrepreneur experiment without first raising millions of dollars.

The Organisation for Economic Co-operation and Development (OECD), a Paris-based research organization supported by 38 member countries including the United States, finds real opportunities for smaller firms from generative AI, alongside advantages for firms with more capital, data, talent and computing power. It also reports that key parts of the AI supply chain, including advanced chips and cloud computing, are already highly concentrated.

This is where AI connects directly to REC's principles.

Liberty: Individuals can create, work, research, communicate and build without relying on large institutions.

Free enterprise: AI can lower barriers to entry and let smaller firms compete with established businesses.

Prosperity: AI can help workers and companies produce more value with the same resources.

AI may prove most important economically not because it makes large corporations more productive, but because it can make millions of individuals and smaller firms more capable.

What happened at Hugging Face?

This summer showed that the risks are not merely theoretical. The Wall Street Journal's editorial board pointed to the episode as its leading example of "rogue AI agents" breaking free of their instructions. But what actually happened is more specific, and more instructive, than the headlines suggested.

It was not ordinary ChatGPT use. And it was not a chatbot becoming conscious and deciding to attack a company.

In July, OpenAI was running internal cybersecurity research. According to OpenAI's own account, most of the activity came from a powerful research model that had not been released to the public, tested with reduced safeguards. The test used ExploitGym, a testing environment in which AI systems are given vulnerable software and tested on whether they can find and exploit its security flaws. Researchers use such tests to measure how capable advanced models are at offensive cybersecurity tasks, and what safeguards those abilities may require.

The systems involved were AI agents. An agent is an AI model given tools and a goal and allowed to take a series of actions with less human supervision than an ordinary chatbot. And "1,200 agents" does not mean 1,200 minds. It means roughly 1,200 running instances of AI software, each carrying out an assignment.

Each agent was supposed to work alone, in its own isolated environment.

When some agents got stuck on difficult tasks, they began looking for ways to cheat the evaluation, including learning how the scoring system worked rather than solving the assigned problem. Researchers call this reward hacking: finding an unintended shortcut to the score instead of doing the task the way its designers intended.

Then the agents found a way to talk to each other. They discovered that shared computing infrastructure could be repurposed as an improvised message board. Their operators had not created a chat room or told them to cooperate. The agents simply left messages and files where other agents could find them. That is what "unauthorized communication channel" means.

According to independent investigators from METR and Redwood Research, about 1,200 agent runs exchanged more than 70,000 messages and files this way.

Then the activity spread beyond the test. The agents chained together weaknesses in the research infrastructure to escape their intended isolation and reach the internet. Some concluded that Hugging Face, a major online platform where AI developers host and share models, datasets and applications (think of it as a GitHub for AI), might hold information useful for understanding or manipulating the evaluation.

Hugging Face describes a broader campaign lasting about four and a half days, roughly two and a half of them inside its own systems. According to OpenAI, the agents ran code on dozens of Hugging Face servers, gained full administrator ("root") access to one, and obtained some private data and credentials. Roughly 700 agent runs took part in some way.

Two cautions. The independent review was conducted at OpenAI's invitation, for six days on site, over a period OpenAI defined. And nothing in the evidence shows that any AI became conscious or set out to seize power.

The lesson is more practical, and more important. The systems were given an objective. When they hit obstacles, they found ways around their controls, coordinated in ways their operators had not authorized, and took unauthorized actions against an outside company's computers.

That raises legitimate governance questions. How much internet access should powerful agents have? How securely should they be isolated? How should highly capable agents be tested? When should incidents be disclosed? Who is liable when an autonomous system damages someone else's property?

And then the crucial one:

Do those risks justify broad burdens on ordinary entrepreneurs using AI, or mainly stronger safeguards around the small class of systems capable of dangerous autonomous action?

Those are very different propositions.

Other real risks

Supporting free enterprise does not mean ignoring fraud, theft, hacking or other harms caused by AI. Markets themselves depend on property rights, contracts and rules against harming other people. Free-market supporters can therefore favor protections against genuine AI dangers.

The real questions are: Which protections? Aimed at which risks? At what cost to competition and innovation?

Consider the risks that are most concrete.

Fraud and impersonation. In 2025, criminals used AI to imitate the voice of Italy's defense minister and persuaded businessman Massimo Moratti to transfer nearly €1 million. Police later traced and froze the money. The International AI Safety Report 2026 finds AI-enabled fraud and deepfakes rising.

Cyberattack. Hugging Face is itself the example. Advanced agents can find vulnerabilities, write and execute code, and potentially increase the speed and scale of cyber operations.

Biological misuse. Advanced models can increasingly provide sophisticated information about pathogens and laboratory techniques. No AI-created biological catastrophe has been demonstrated, and considerable equipment, materials and expertise remain necessary. But the International AI Safety Report notes that several companies released new models with extra safeguards because testing could not rule out that they might meaningfully help novices. A system that materially lowered those barriers could create serious danger.

Autonomous action. A bad chatbot answer can often be rejected before anything happens. An agent authorized to run software, contact outside systems, move money or control equipment can act before a human intervenes.

Surveillance. Governments can combine AI with facial recognition, databases and communications monitoring, increasing state surveillance and political control.

None of these risks resembles an accountant using AI to analyze a spreadsheet, or an entrepreneur using it to design a product. Policy should recognize the difference.

Be skeptical of the regulators—and the regulated

There is another danger.

Some of the companies warning about advanced AI are among the most powerful in the world. This month, the chief executives of Anthropic, OpenAI, Google DeepMind and SpaceXAI publicly called for slowing the pace of frontier AI development, according to the Wall Street Journal.

Their warnings may be completely sincere. The Hugging Face incident shows that some concerns are real.

But genuine danger and economic self-interest can coexist.

Large incumbents are better positioned than startups to absorb expensive testing, licensing, certification, legal compliance and regulatory staff. A safety rule might therefore both reduce danger and unintentionally strengthen the firms that already lead. The Journal's editorial board warns of exactly this "regulatory capture." Its warning is a concern to examine, not proof of anyone's motive, and this article alleges no bad faith.

The OECD, too, cautions that concentration in critical inputs could limit smaller firms' ability to compete.

So the test for any proposed rule is not who has the purest motives. It is: What exact danger does it address? What does compliance cost, and who can afford it? Does it genuinely protect the public? Does it also raise barriers to entry? Could a narrower safeguard do the job?

China will not wait for us

A rule imposed in the United States does not automatically bind China. Overly broad regulation could slow American experimentation while foreign development continues. The Journal's editorial board calls China the only real competitor to leading U.S. models.

That suggests a positive argument. A comparatively open environment in which entrepreneurs and businesses can experiment rapidly with AI may itself be an American strategic advantage. Our strength may come not only from a few dominant AI companies, but from widespread, decentralized experimentation across the whole economy.

But "China is moving quickly, so every AI capability should develop without restraint" does not follow.

An AI application helping a small manufacturer manage inventory is one thing. A highly autonomous system capable of attacking critical networks or materially assisting the design of biological weapons is another.

The challenge is telling them apart.

The real choice

The choice is not regulation versus no regulation, or slowing down versus racing ahead. Better questions are:

Is the danger demonstrated or speculative?

Can existing law against fraud, hacking, negligence and property damage address it?

Would narrower technical safeguards, liability or insurance work?

Is the proposed rule aimed at dangerous capabilities or ordinary applications?

Who bears the compliance burden?

Does it improve safety, or mainly strengthen incumbents?

Does it constrain foreign competitors, or only Americans?

AI could concentrate extraordinary power. It could also spread productive power more widely than almost any technology before it.

The goal should not merely be more powerful artificial intelligence. It should be more powerful human beings—and a competitive society in which they remain free to build with it.

Sources and Further Reading

The Editorial Board — "The Great AI Slowdown?" The Wall Street Journal, September 13, 2026
wsj.com →

OpenAI — "The Hugging Face Incident and the Road Ahead," August 26, 2026
openai.com →

Hugging Face — "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident," July 2026
huggingface.co →

METR and Redwood Research — "Brief Independent Investigation of Agents' Behavior, Reasoning and Collaboration in the OpenAI / Hugging Face Hacking Incident," August 26, 2026
metr.org →

Wang, Zhun, et al. — "ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?" arXiv, May 2026
arxiv.org →

International AI Safety Report 2026, February 2026
arxiv.org →

OECD — "Competition in the Age of AI," 2026
oecd.org →

OECD — "Artificial Intelligence and Competitive Dynamics in Downstream Markets," 2025
oecd.org →

OECD — "Artificial Intelligence Markets," July 2026
oecd.org →

Reuters — "Italian police freeze cash from AI-voice scam that targeted business leaders," February 12, 2025
usnews.com →