Anthropic's 'AI worm' warning shows the weak point is billions of cheap, unpatched home devices, not the model itself

Anthropic published a report on Monday warning of "AI worms": self-replicating AI agents that could break into Wi-Fi-connected TVs, cameras and other household devices and use them as a base to spread further. A worm is malware that copies itself to other machines by scanning the network it is on. According to the report, this could be only months away if models keep improving at the current rate. "We expect a future awash with agents running on all types of hardware, including some that aren't under human control," the company writes.

The experiment behind the warning is modest but concrete. Anthropic's researchers ran a lightweight AI model across 15 smart TV boxes with just 4 GB of RAM each, about the memory of a budget phone. The report cites earlier research showing that a worm could carry its own model weights (the trained parameters that make up the model) and start running inference, meaning the model actually executing, on hardware it takes over. On weak devices it would spread slowly, which the report describes as a kind of "digital slime mold." The most important finding is qualitative: unlike conventional malware, these agents could work out unfamiliar devices with no outside help.

The report also makes an economic argument. It calls datacenter GPUs "a clear, high-value target," but notes that Internet of Things (IoT) devices, "cheap and typically poorly defended," could together provide enough compute to run a capable model. Smart fridges, robot vacuums, light bulbs and cameras from little-known manufacturers have been insecure for years. What changes now is who does the attacking. "Lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world's IoT compute," the report says. Human attackers skip low-value targets because their time is limited. An agent has no such limit.

A caveat on how the story is being framed. Futurism, which reported on the findings, describes Anthropic as "no stranger to dramatic and doomy proclamations," and the report itself is careful: "We don't yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain." What has actually been shown is that pieces of the chain work, such as running a model on weak hardware and adapting to new devices. A self-sustaining worm spreading on its own has not been shown, and "months away" is a projection that depends on current trends holding. Treating it as an existing threat would mean regulating the panic rather than the real capability.

Our take: this fits our view that AI's real danger is short-term and practical. The near-term risk is the automation of old attacks, not a distant superintelligence. Agentic AI is already industrialising fraud, and this report extends that logic to hardware. The weak link is not the frontier model. It is decades of IoT built with no security incentive, which agents can now exploit at near-zero marginal cost. That puts the burden on manufacturers and regulators: mandatory updates, no default passwords, and network segmentation in home routers. The same principle we see in regulating agents applies here too: what matters is what an agent is allowed to access and run, not only what the model says. The longer-term upside is that the agents able to map unfamiliar devices are also the defenders able to audit and patch billions of them, a job that was never economic for humans. If the industry treats this as a demonstrated risk to fix now, rather than a headline to dismiss or exaggerate, it is a solvable problem.