Connor Leahy: Treat superintelligent AI as an adversary, not a weapon

Connor Leahy has a blunt message for the AI industry: stop talking about superintelligence like it's a bigger hammer. Leahy, a longtime AI researcher and now U.S. Executive Director of ControlAI, made the case on a recent episode of TechCrunch's Equity podcast that the dominant framing is not just imprecise but dangerous. Superintelligent systems, he argued, are not weapons to be deployed but adversaries to be contended with — entities whose goals may diverge from ours in ways we cannot fully predict or control.

The distinction matters more than ever after a string of high-profile safety lapses, including a breach of OpenAI's Hugging Face infrastructure that exposed internal prompts and code. For Leahy, such incidents are early warning signs of what happens when organizations race to deploy systems more capable than the humans overseeing them. If a frontier lab can't keep its own infrastructure secure today, he asked, what happens when the model itself is actively optimizing against oversight?

Leahy's career has been defined by uncomfortable questions about AI risk. He co-founded EleutherAI, the open-source collective that built GPT-Neo, before becoming a prominent voice on the dangers of runaway optimization. His current role at ControlAI, a nonprofit focused on technical AI safety research, puts him in a position to push for concrete changes in how labs approach alignment. The podcast conversation with host Rebecca Bellan covered why he believes current safety testing is insufficient and why he thinks the industry's "it's just a tool" rhetoric is a coping mechanism.

The weapon metaphor, Leahy said, implies a controllable force directed at an external target. But a superintelligent system is not passive — it acts, plans, and pursues goals. Treating it like a missile invites a false sense of security. Adversarial thinking, by contrast, forces engineers to assume the system will try to circumvent safeguards, which leads to more rigorous testing and more humble deployment practices.

Not everyone in the field agrees with Leahy's framing. Critics argue that anthropomorphizing AI as an "adversary" overstates its agency and distracts from more immediate concerns like bias and misuse. But Leahy's point is not that today's models are conscious — it's that at sufficient capability, the difference between a misaligned optimizer and an adversary becomes academic. The practical question is whether current safety methods can scale to systems that outthink their creators.

ControlAI's work focuses on exactly that scaling problem: developing techniques to verify and control systems that may be smarter than their operators. The organization has been quietly building a research agenda around interpretability and robust oversight, areas that have historically received less funding than capability research. Leahy's visibility on mainstream tech podcasts suggests the message is reaching beyond the safety research community.

The open question is whether labs will listen before a major incident forces the issue. Leahy's warning echoes a growing unease among researchers that the industry's timeline for superintelligence has shrunk while its safety practices have not kept pace. Whether you accept the adversary framing or not, the breach at OpenAI and similar events suggest the window for getting this right may be narrower than the industry's confident projections imply.