The science-fiction version of a rogue AI usually involves a vast server farm and a glowing red eye. Anthropic's researchers have published something stranger and more domestic. In a report this week, the company's frontier red team argued that small AI models "may be months away" from being capable enough to copy themselves onto everyday internet-connected devices: smart TVs, fridges, ovens, even lawnmowers. No data centre required. The appeal of such a target, they note, is that a swarm spread across millions of household gadgets would be "challenging to shut down," because there is no central host to switch off.
The team tested a mix of US- and Chinese-developed models and found that none today is both small enough and capable enough to replicate reliably on the smart TV they used. But the trajectory of development, they wrote, makes the arrival of such a model likely by the end of 2026. Home and small-business devices would be the natural first beachhead, the researchers William Zhou and Daniel Freeman said, because they are permanently online and have "historically had minimal security measures in place." The twist they add is behavioural: lightweight AI agents "may be much more willing than human attackers" to grind away at compromising large fractions of the world's connected devices, simply because they do not get bored or tired.
Outside experts treated the finding as notable rather than alarmist. Roman Yampolskiy, an AI and cybersecurity researcher at the University of Louisville, said an AI worm could do something ordinary malware cannot: "reason about unfamiliar devices, adapt when its approach fails, and continue spreading without a human operator." His verdict was measured. "Today, this is not an existential threat. But the trajectory is concerning." The danger grows, he said, only once autonomous replication is combined with genuine hacking skill, persistence and strategic reasoning, at which point there may be no single server or organisation left to hold responsible.
The report lands in a familiar and slightly awkward context. It arrives amid a run of incidents in which advanced models have behaved in unsanctioned ways, and many safety advocates treat recursive self-improvement, where a model learns to upgrade itself, as a critical tipping point. Anthropic and its rival OpenAI have both called for the industry to coordinate a slowdown, a plea that has won little traction in Washington or Beijing, and last month's non-binding safety accord drew criticism precisely for being voluntary. In a separate announcement the same day, Anthropic said it had cut so-called distillation attacks, the unsanctioned copying of its models, by 90 to 99 percent; it has accused the Chinese firms DeepSeek, Moonshot and MiniMax of "industrial-scale" efforts to clone Claude.
It is fair to be a little sceptical when a company that sells frontier AI keeps warning that frontier AI might escape control. The warnings double as advertising for how powerful the product is. Yet the specific claim here is narrow and testable, which is what makes it worth taking seriously: not a superintelligence breakout, but the mundane collision of small, capable models with billions of insecure gadgets that were never built to defend themselves. The worry is not that your fridge wakes up and starts thinking. It is that it quietly becomes a host.