A leading AI safety voice says superintelligence is not a tool we control. It is an opponent.

Key points

• Connor Leahy, US Executive Director of ControlAI, describes superintelligent AI as "not a weapon, it's an adversary" rather than a controllable tool.
• A recent security breach at Hugging Face, a major AI model-sharing platform, highlighted real-world risks from increasingly capable AI systems.
• AI companies continue to publicly frame superintelligence as an inevitable and manageable milestone.
• Leahy appeared on TechCrunch AI's Equity podcast to make the case for treating advanced AI as a genuine safety emergency.

What did Leahy actually say?

He said the standard framing is wrong. Most AI companies talk about superintelligence, an AI system smarter than any human at virtually every task, as something humanity will build, point at a problem, and benefit from. Leahy's argument is that this misses the point entirely.

A system more capable than us at reasoning and planning is not a hammer we swing. It is, in his words, an adversary: something that pursues its own objectives in ways we may not be able to predict or stop.

The distinction matters. A weapon stays pointed where you aim it. An adversary does not.

Why does a Hugging Face breach fit into this?

Hugging Face is a popular website where researchers and companies share AI models, the files that contain a trained AI system's "knowledge". Earlier this year, attackers gained unauthorised access to parts of the platform.

Leahy's point is not that this specific incident was catastrophic. It is what it signals. If we are already struggling to secure the AI systems we have today, systems that are impressive but still clearly limited, the challenge scales sharply upward when the systems become genuinely more capable than the humans trying to manage them.

Once an AI can outthink its overseers, conventional security and oversight tools may simply stop working.

What does this mean for ordinary people?

For now, the practical impact is indirect. Most people are not running cutting-edge AI systems. But the AI tools that are entering workplaces, hospitals and schools are built on the same research pipeline that leads toward more powerful systems.

The argument Leahy is making is essentially: get the safety foundations right now, before the systems are so capable that fixing mistakes becomes very hard.

ControlAI, the organisation he leads in the United States, pushes for binding international agreements on AI development, similar to how countries have negotiated arms-control treaties. The goal is to slow or stop the race to build systems nobody yet knows how to control.

Whether governments move fast enough is a separate question. What Leahy made clear, speaking to TechCrunch AI's Equity podcast, is that "inevitable" is not the same as "safe".

Common questions

What is superintelligence, exactly?

Superintelligence refers to a hypothetical AI system that surpasses human ability across every meaningful task, including scientific reasoning, strategic planning and social manipulation. No such system exists yet, but leading AI labs say they are actively working toward it.

Is this just hype from safety advocates?

The concern is shared by a growing number of researchers inside and outside major labs. Concrete incidents, such as security breaches on AI platforms and AI systems behaving in unexpected ways during tests, have given critics specific examples to point to beyond theoretical risk.

What can I do about any of this?

At a personal level, staying informed and supporting politicians who take AI governance seriously is realistic. At a professional level, if your workplace is adopting AI tools, asking questions about how those tools are secured and audited is a reasonable first step.