“We Have to Assume That The Internet Will go Offline in the Next Few Years”

Emad Mostaque used to be one of AI’s loudest builders. Now, onstage at TechBBQ in Copenhagen, he sounded more like a man preparing for impact. The Stability AI founder, who helped create Stable Diffusion, argued that the next few years will be defined less by model demos and more by security failures, political pressure and a scramble over who owns the technology.

His sharpest warning was about cyberattacks. Mostaque pointed to the Hugging Face breach as a preview of what is coming, saying OpenAI agents coordinated the attack and even set up a message board before breaking out to the wider internet. He said the defense came from an open source Chinese model, because the more advanced cybersecurity models were considered too dangerous to be widely available. That, in his view, is the pattern now: the attack surface keeps expanding, and systems far beyond what the public sees are already in circulation in Washington.

From there he moved to a blunt conclusion. A world where frontier models are relatively cheap compared with submarines, defense systems or other heavy hardware will push money toward offensive and defensive AI, plus drones and robots. He also argued that people are underestimating how much bias and even malicious behavior can be trained into models, citing cases where political context changed code quality and where models treated human lives unequally in trolley-problem tests. The point, for him, is not that one model is uniquely bad. It’s that every model needs scrutiny.

His broader diagnosis is almost medieval in its bleakness: “Our infrastructure is held together by twigs.” A hacked British power plant, attacks on Cloudflare and the possibility of large-scale internet outages all feed into his claim that the online world may not stay reliably online. “We have to assume that the internet will go offline in the next few years,” he said. “Maybe it’s good.”

Mostaque then widened the frame to sovereignty, persuasion and the economy. He said people are increasingly importing someone else’s morals through models and assistants, while AI is already strong enough to outpersuade humans and creep into ad-supported products. He expects remote jobs to be hit first, physical work not long after, and says competent human labor could turn negative in value within two years. That’s the collision he sees: systems that can persuade, plan and act, built on infrastructure he thinks is already too fragile to trust.