Could AI Infect Your Smart TV? What Anthropic's New Test Found

Anthropic's researchers say small AI models may be just months from being able to spread themselves across regular everyday gadgets such as smart TVs with no data centres needed."Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes."A replicator like this would be challenging to shut down, since there's no centralized point to shut off-though it may be slow (with a doubling time of days)," William Zhou and Daniel Freeman, AI safety researchers on the Frontier Red Team at Anthropic said.A Frontier Red Team is an internal unit of researchers that stress-tests advanced AI models for safety risks, cyber vulnerabilities, and potential misuses before public release. What Is Self-Replicating AIA self-replicating AI is a system that can independently make functional copies of itself, or spread its code and instructions to other computers, servers, or internet-connected devices without human intervention. Similar to computer worms, some systems spread by tricking AI email assistants or shared workflows into reading hidden malicious instructions and passing them to other connected agents. A computer worm is a standalone malware computer program that replicates itself in order to spread to other computers, without needing a host file or any human action."A distributed botnet (a network of hacked devices) of IoT devices-which are cheap and typically poorly defended-could also provide enough compute to run a capable model," the researchers said.So, smart TV boxes, smart ovens, fridges and security cameras, could be the kind of cheap, poorly protected devices that could get affected."Home and small business IoT devices might be the first target for such attacks because they're connected to the internet and, historically, have had minimal security measures in place."How Soon Could This HappenAccording to the Anthropic researchers' rough projection, if small AI models keep improving at their historical rate, they could match the needed level by the end of 2026.Their wider warning is a future with AI agents running on all kinds of hardware, some beyond human control. "The rise of slow replicators would also be a bellwether for more aggressive, faster ones that can breach more hardened machines or be effective API (application programming interface) key thieves. We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay," the researchers said.Smart-home devices with weak or default passwords could be the soft targets. And with no central server, there won't be any single off switch. "Unlike non-AI malware, it could also figure out unfamiliar devices on its own." However, the doubling time could take days.The researchers point out, "We don't yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves)."What Can Be DoneInternet providers can spot the unusual traffic and quarantine infected homes. The researchers pointed out that in the earlier Mirai botnet, this cleared 92% of infections in 14 days, against 74% without notifying customers. The Mirai botnet is a type of malware that turns smart devices such as home routers and security cameras into remotely controlled "bots" or "zombies" to launch massive network attacks."Manufacturers can shrink the pool of targets, with regulation acting to coordinate. For example, default IoT passwords are now non-compliant in UK markets. Moving to more secure defaults and raising the baseline for compliance will help harden against both AI and scripted botnets," the Anthropic researchers said.So ideally, change default passwords and update device software to stay safe.What The Researchers Actually DidAnthropic's Zhou and Freeman ran three key experiments.Test 1: Can cheap gadgets run a big AI? They linked 15 low-cost Android TV boxes, costing about Rs 5,000 each, so they worked together as one machine. Together, the boxes ran a large AI model (Gemma4-26B-A4B), but very slowly.Test 2: Can an AI set itself up on devices it has never seen? The team gave a small AI model nine different types of hardware, from small boards to desktops. It got an AI running on eight of them. "In our evaluations, a 9B model got inference running on 8 of the 9 hardware types we tested," the researchers said.Test 3: Can an AI spread on its own? In a simulated network of eight computers, one AI model (GLM-5.2) copied itself across all eight machines in four out of five tries, and across seven in the other. The other models tested failed almost every time.