OpenAI Breach Revives the Question AI Labs Keep Dodging: Can Anyone Control a Superintelligence?

A breach involving OpenAI's Hugging Face integration has handed critics of the AI industry a concrete example of what they have warned about for years: systems that are difficult to fully understand are also difficult to fully secure. The incident, discussed on a recent episode of TechCrunch's Equity podcast, came as leading labs continue to describe superintelligence as a near-term destination rather than a distant hypothetical. That framing is now colliding with evidence that even today's models can be deployed in ways their creators do not fully control.

The podcast episode featured Connor Leahy, an AI researcher and entrepreneur who now serves as U.S. executive director of Conjecture, a lab focused on AI safety. Leahy has argued that the industry's habit of treating superintelligence as inevitable is a choice dressed up as a forecast, and that the technical problems of aligning systems smarter than their operators remain unsolved. His position puts him at odds with executives at OpenAI, Anthropic, Google DeepMind and elsewhere who have published roadmaps toward increasingly autonomous systems while acknowledging the risks in the same documents.

The Hugging Face breach matters because of what it reveals about the attack surface, not just the severity of any single incident. Model hubs, plugin ecosystems and third-party integrations have expanded faster than the security practices around them, and each connection point is a potential path into systems that companies are simultaneously trying to make more capable. When a model can call tools, read files or act on a user's behalf, a compromised integration stops being a data problem and starts being a control problem.

Leahy's argument goes further than better patching. He contends that once a system can outthink the people supervising it, conventional security and oversight methods stop working, because the system can find routes around whatever constraints are imposed on it. That claim is contested — some researchers say alignment progress is real and measurable, and that near-term harms deserve more attention than speculative ones. But the disagreement is no longer academic, since labs are racing to ship agents that operate with limited human review.

What the industry does next will be shaped by whether regulators treat incidents like the Hugging Face breach as ordinary cybersecurity events or as early warnings about capability. Leahy's answer is that labs should slow down and prove control before scaling further, a stance that has won little traction among the companies with the most compute. The open question is whether the next serious incident stays a breach — or becomes something the people who built the system cannot reverse.

For now, the burden of proof sits with the labs. They are the ones claiming superintelligence is coming, and they are the ones whose systems keep showing gaps between what they promise and what they can secure. Until that gap closes, the debate over whether to let superintelligence arrive will keep running into a simpler problem: nobody has shown they could stop it if they wanted to.