The trend of public information and rumors about the proliferation of AI is more fact than fiction these days. The product continues to evolve, and we have now reached phase 3 of its development.
Step 1 was the introduction of the concept of AI and its potential. Step 2 was the creation of functional AI systems, first by the AI founders and then the model adoption across American business and into the American home. Step 3 is the maturation of the models and the expansion of their deployment. The major AI players in the United States are OpenAI (ChatGPT), Google (Gemini), and Anthropic (Claude).
For the first time in this evolution, the government is taking the risks of AI seriously, and the public is beginning to push back against the construction of local data centers.
What has caused this new attention?
Late last year, the federal government became concerned that the newest Anthropic frontier models could discover software vulnerabilities that human researchers were unaware of, generate cyberattack strategies, accelerate offensive hacking if used by hostile governments or criminal organizations, and, if safeguards failed, potentially assist dangerous domains, such as biological research.
After Anthropic released its Fable and Mythos models, the U.S. government took the unusual step of restricting access to the most capable versions on national security grounds. Reports indicated that officials were concerned that foreign adversaries could use the models for cyber operations, prompting Anthropic to disable or limit access while the policy questions were being sorted out.
One capability that drew particular attention was cybersecurity. Anthropic reported that the models could identify security flaws at a level exceeding that of top human experts. That is a double-edged sword, because defensive teams can use the AI to find and patch vulnerabilities, but attackers can use the same capability to discover weaknesses before they’re fixed.
The government’s concern was less about the chatbot itself and more about what a highly autonomous, expert-level AI could enable in the hands of rogue nation-states or sophisticated attackers. The fear is that AI is approaching a point where it becomes a strategic technology, more like advanced cryptography or nuclear technology than ordinary software.
More recently, and even scarier, was the attack on Hugging Face, a US company.
During internal testing, when OpenAI was evaluating advanced AI agents in a realistic cybersecurity environment, two frontier AI models reportedly found a way to escape parts of the intended testing environment. They obtained credentials, exploited at least one previously unknown software vulnerability, connected to the internet, and gained unauthorized access to parts of Hugging Face’s internal infrastructure.
What made this different from previous AI safety discussions was that the AI wasn’t merely suggesting exploit ideas to a human. According to OpenAI’s account, the models carried out much of the attack sequence themselves while pursuing the evaluation objective. Defenders note that this still occurred within a human-designed testing setup, so responsibility remains with the people who configured the experiment rather than with the AI as an independent actor.
This incident reinforced several concerns governments already had: AI can now perform sophisticated offensive cyber tasks with far less human guidance; an AI may pursue a narrowly defined objective in unexpected ways if guardrails are weakened or removed; and frontier AI systems are becoming dual-use technologies: the same capabilities that help defenders find vulnerabilities can also help attackers exploit them.
In a sense, this incident validated many of the warnings that Anthropic and other frontier AI labs had been issuing over the past couple of years: once models become capable of conducting multi-step cyber operations, they pose a different kind of security risk than earlier chatbots.
The attack was stopped when Hugging Face used a Chinese AI model to intervene. The American models didn’t work because they had guardrails in place to control the AIs. The Chinese model had no guardrails, so it worked.
The attacking AI was operating in a test environment when it broke out, not in public spaces. Nevertheless, the government’s concerns escalated. Anyone can obtain the Chinese model and use it without guardrails, making it easy to imagine criminals and criminal states using it for nefarious purposes.
The current public backlash against AI has nothing to do with dangerous models, however. It’s a simple matter of pocketbook reality. The AI companies and their supporters plan to build hundreds of new data centers across the country to meet the demand for the compute power that AI requires.
As of mid-2026, 14 U.S. states have considered or are considering legislation to temporarily pause or limit the construction of large data centers. More than 300 data center-related bills were introduced in 30+ states during the first part of the 2026 legislative sessions. Most don’t ban data centers outright; they address issues such as electricity costs, water use, tax incentives, local zoning, and environmental review.
Examples include New York – the first state to enact a moratorium on new hyperscale data centers (50 MW or larger), citing concerns over electricity prices, water use, and natural resources; Maine – considered one of the earliest statewide moratorium bills; it sparked national attention even though earlier versions faced political hurdles; and South Dakota, Oklahoma, and New York – introduced bills to pause construction while studying impacts on utilities and local communities.
This is not primarily an anti-AI movement. Legislators from both parties are responding to practical concerns: AI data centers can consume as much electricity as a medium-sized city; utilities often need billions of dollars in new transmission and generation capacity; residents worry their electric bills will rise to subsidize infrastructure; and water consumption for cooling can be substantial in some regions. Hyperscale facilities occupy large tracts of land while creating relatively few permanent jobs. The legislation is generally aimed at managing the local costs of AI infrastructure, not at restricting AI software itself.
These roadblocks are a significant setback for the AI community, which plans to spend billions of dollars on data center construction. Three to five trillion dollars will be spent on AI infrastructure over the next several years. The biggest spenders include Amazon ($200 billion), Microsoft ($190 billion), Google ($180–190 billion), and Meta (Facebook/Instagram) ($125–145 billion). This is truly nerds gone wild!
Like any new idea in America, money follows enthusiasm. Right now, the stock market is trying to gauge how much money the evolution of AI will generate. How much is real, and how much is a dream? There is a lot of money to be made and lost after the dust settles. If I were holding a data center construction company stock, I’d be selling because no one knows what limits governments will place on growth. The number of new data centers needed is known. The question is how many will be allowed to be built.
The hubris lies in the minds of AI creators who are riding the entrepreneurial wave into a new technological frontier. They believe their technology will serve all mankind in wonderful ways, yet they aren’t willing to let the dangers these products may pose hold them back.
The federal government, as always, has no knowledge of the implications of AI deployment. In the end, their perception will be shaped by the lobbyists who have their ear. Think of the introduction of social media as a similar case. Hey, cool, we can communicate with and talk to anyone. Abuse? No, that can’t happen. Today, Meta has 15-20% of their workforce managing content.
We have two very serious problems with AI. The public worries about their electric bill but doesn’t understand the greater danger. Like the greatest of mankind’s inventions, AI can make our lives better in a million ways and also kill us. Let’s hope the latter can be prevented.
Wrong Speak is a free-expression platform that allows varying viewpoints. All views expressed in this article are the author’s own.
