Who Really Controls AI? Inside the Global Power Struggle Over Advanced Models

Powerful AI systems are moving from science fiction to real-world infrastructure faster than most people can follow. The latest flashpoint is a model called Mythos, described as so capable and so dangerous that its creators say it can’t be released to the public. That claim has triggered emergency meetings in governments and banks, and a bigger question: who is actually in control of AI?

What Is Mythos and Why Is It So Controversial?

Mythos is a cutting-edge AI model developed by Anthropic, one of the leading AI labs in the United States. The company says Mythos is powerful enough to automatically find and exploit cybersecurity flaws in the software that runs banking systems, utilities, power grids, and other critical infrastructure.

In the right hands, Mythos could help discover and fix vulnerabilities before attackers find them. In the wrong hands, it could enable unprecedented cyber attacks at global scale. That dual use is what makes it so controversial.

Anthropic has decided not to offer Mythos as a public product. Instead, it’s running a restricted program with a few dozen of the world’s largest tech and financial institutions, including Apple, Amazon, and JPMorgan Chase. The idea is to give these players a head start in patching their systems before similar models spread more widely.

At the same time, Anthropic has acknowledged reports of unauthorized access to Mythos, even within this controlled setup. There’s no evidence of malicious use so far, but the incident has sharpened concerns about how secure these ultra-powerful models really are.

Agentic AI: From Chatbots to Autonomous Hackers

Mythos isn’t just another chatbot. It represents a broader shift toward what experts call “agentic AI” – systems that don’t just respond to prompts, but can take actions, plan, and operate with a degree of autonomy.

In simple terms, imagine the difference between:

• A regular AI assistant that answers questions and writes emails when you ask it to.
• An agentic AI that can browse systems, run tools, execute code, and keep working toward a goal with minimal human supervision.

Applied to cybersecurity, that means one skilled attacker could effectively control the equivalent of thousands of tireless, automated hackers. Those agents could probe banks, telecom networks, power companies, and government systems at machine speed.

The same technology can also be used defensively. Companies can deploy AI agents to constantly scan for vulnerabilities, monitor networks, and respond to threats faster than human teams ever could. But this sets up an arms race: offensive and defensive AI agents escalating against each other.

If you want a deeper dive into how enterprises are starting to secure their AI stacks and infrastructure, it’s worth reading this guide to enterprise AI security, tools, and access control.

Is AI Already Out of Control?

Many experts argue that AI development has already slipped beyond meaningful public oversight. A few key concerns keep coming up:

1. Opaque systems and unknown data

Most people have no idea what data was used to train these models, how they actually work under the hood, or how decisions are made. Creators, artists, and everyday users often discover their data has been used only after the fact.

2. Concentrated power in a few companies

The most advanced models are controlled by a small cluster of US-based giants and a handful of Chinese players. Decisions about safety, access, and deployment are largely made in corporate boardrooms, not parliaments or public forums.

3. Real-world impacts on jobs, costs, and the environment

AI data centers are driving up local energy and water use, while automation is reshaping job markets. Some estimates suggest that within a decade, AI infrastructure could consume as much energy as a large country. People are starting to ask: who benefits, and who pays the price?

4. Social media déjà vu

There’s a strong sense of “we’ve been here before.” Social media platforms were allowed to scale globally with minimal guardrails, and only later did societies grapple with misinformation, polarization, and privacy abuses. Many fear AI is following the same path—only faster and with higher stakes.

How Different Regions Are Trying to Control AI

There is no single global AI rulebook. Instead, three main power centers are emerging, each with its own approach: the US, China, and the European Union.

The US: Innovation First, Regulation Later

The US is home to most of the leading AI labs and cloud providers. The current political stance from the Trump administration is clear: avoid regulation that might slow American dominance in AI. Even attempts by individual US states, like California, to introduce their own AI rules are facing pushback from Washington.

Behind the scenes, big tech companies are some of the most powerful lobbyists in US politics. That makes it hard to imagine strong, binding AI regulations emerging quickly at the federal level. Instead, the US is relying heavily on voluntary commitments and self-regulation from the very companies racing to outdo each other.

China: Tight Control and Strategic Expansion

China has taken a much more hands-on regulatory approach. Authorities require AI platforms and large language models to register, and they regulate both the “pipes” (infrastructure) and the “platforms” (apps and services).

At the same time, Chinese companies are aggressively open-sourcing their models and pushing them into the Global South. The strategy is similar to what we’ve seen in telecom infrastructure: offer low-cost or free technology to build long-term influence in emerging markets.

China has even proposed a World Artificial Intelligence Cooperation Organization, headquartered in Shanghai, as a kind of global AI coordination body—an answer to Western-led initiatives.

Europe: Strong Rules, Weak AI Industry

Europe is about to roll out the world’s first comprehensive AI law, the EU AI Act, which comes into force in stages starting this year. It focuses on risk categories, transparency, and user protections, and aims to prevent the worst abuses before they spread.

The catch: Europe doesn’t have many frontier AI labs of its own. It’s a bit like building an elaborate signaling system for train tracks when you don’t actually run many trains. The heavy compliance burden is relatively easy for US and Chinese giants to absorb, but can be crippling for European startups and mid-sized companies.

That raises a tough question: will Europe’s rules mainly end up regulating foreign tech, or will they also unintentionally hold back its own AI ecosystem?

Global Governance: Can We Treat AI Like Nuclear Tech?

Some policymakers and researchers argue that AI should be governed like nuclear weapons and nuclear energy: with global treaties, inspections, and an international watchdog agency.

In theory, that could mean:

• Shared safety standards for the most powerful models
• International inspections of data centers and training runs
• Limits on exporting certain AI capabilities to unstable regions

In practice, there are big obstacles:

• Major powers don’t agree on what should be restricted or who should enforce rules.
• Past global agreements, like the Paris climate accord, show how hard it is to turn non-binding promises into real action.
• Some states simply ignore or sidestep UN processes when it suits them.

Most experts in the discussion agree a classic “non-proliferation” treaty for AI is unlikely to work on its own. AI is easier to copy and distribute than nuclear material, and the incentives to develop it are enormous.

Emerging Ideas: Sovereign AI and Open Source

With a US–China duopoly looming, other countries and regions are looking for ways to avoid total dependence on a few foreign tech giants. Two ideas keep coming up: sovereign AI and open source.

Sovereign AI

Sovereign AI is the idea that each country (or region) should control its own AI infrastructure, data, and key models. Instead of relying entirely on US or Chinese cloud providers, governments and large enterprises invest in local data centers, chips, and models tailored to their language, laws, and culture.

We’re already seeing moves in this direction in places like Europe, India, the Middle East, and Japan. The goal is not total isolation, but reducing strategic dependence on a handful of foreign vendors.

Open Source AI

Open source AI models make their code—and sometimes their training methods—public. That lets universities, startups, and smaller countries build on top of them without paying huge licensing fees or being locked into a single vendor.

Supporters argue that open source:

• Spreads AI benefits more widely
• Allows more independent safety research and auditing
• Reduces the power of any single company or country to dictate terms

Critics worry that open-sourcing very powerful models could also make it easier for bad actors to weaponize them. The debate is far from settled, but many see open source as one of the few realistic ways to give the rest of the world a seat at the AI table.

If you’re interested in how one of the other frontier models is evolving, and how that ties into this broader competition, take a look at this breakdown of Grok 4.3 and xAI’s strategy.

What Needs to Happen Next?

There’s no single fix for AI governance, but several priorities are emerging across experts and policymakers:

1. Stronger cybersecurity and AI safety practices

As models like Mythos appear, both governments and companies will have to invest heavily in AI-driven defense, robust access controls, and independent safety evaluations.

2. Democratic accountability

Decisions about how AI is used shouldn’t be left entirely to a handful of CEOs. Legislatures, regulators, and civil society groups need real visibility into how these systems are trained, deployed, and monitored.

3. More global voices at the table

The current AI race is dominated by the US and China, with Europe trying to regulate from the sidelines. Countries in the Global South risk becoming passive consumers of whatever technology they’re given. Sovereign AI projects and open source ecosystems are two ways to rebalance that.

4. Focus on real-world harms, not just sci‑fi scenarios

Apocalyptic visions of AI “doomsday machines” grab headlines, but the most urgent issues are often more mundane: job displacement, energy use, surveillance, cyber risk, and loss of privacy. Those are the areas where regulation and public debate can make a difference right now.

Mythos is a warning shot. It shows how quickly AI capabilities are advancing, and how much power is concentrating in a few hands. The next few years will likely decide whether AI becomes a tool that broadly benefits societies—or a technology that deepens inequality, insecurity, and geopolitical tension.