Anthropic’s Self-Replicating AI Research Is an IoT Security Story

Anthropic's Frontier Red Team says small AI models may be months from copying themselves across cheap IoT devices, using $55 Android TV boxes as the test bench. Here's what they found, why IoT is the first target, and how to rank connected devices by the compute they carry, not just the data they hold.

IoTPortal Security graphic: playful illustration of cheap TV streaming boxes on a shelf passing a glowing AI spark from one to the next while a calm cellular router with a shield badge watches, headlined Anthropic's AI worm research is an IoT story
Quick answer: Anthropic's Frontier Red Team says small AI models may be months from copying themselves across cheap IoT devices, using $55 Android TV boxes as the test bench. Here's what they found, why IoT is the first target, and how to rank connected devices by the compute they carry, not just the data they hold.

Anthropic’s self-replicating AI research has been reported as a story about AI safety. Read it properly and it’s an IoT security story. On 12 October, Anthropic’s Frontier Red Team published AI models may be capable of self-replicating without access to a datacenter. Their conclusion is that small AI models may be months away from copying themselves across cheap, badly secured connected devices, with no data centre involved. Their test bench wasn’t a rack of GPUs. It was fifteen off-brand Android TV boxes at about $55 each. I’ve spent 25 years putting connected kit into places nobody looks at again, and this one made me sit up. It changes the question we ask about the cheap boxes on our networks. We used to ask “what data can it leak?” Now we also have to ask “what could it run?”

The short answer: Anthropic showed that a small open-weight AI model can be split across cheap IoT devices and run there, and that some models can already set themselves up and spread across a simplified test network on their own. Nothing like this has been seen in the wild. But the researchers expect small models to reach that capability level by the end of 2026, and they name home and small-business IoT as the likely first target because it’s internet-connected and poorly secured. For anyone running connected estates, the defences are the boring ones: no default passwords, no open debug ports, private networks and someone watching the traffic.

What Anthropic actually tested

The researchers, William Zhou and Daniel Freeman, break autonomous replication into four stages: find a target, get code running on it, set up AI inference on the new host, then carry on spreading from there. They deliberately skipped the hacking stages, which earlier research has already covered, and focused on the bit nobody had measured: can a small model get itself running on hardware it has never seen?

ExperimentWhat happened
Inference setup on unfamiliar kitA 9-billion-parameter model got inference running on 8 of 9 hardware types on its own, from Nvidia Jetson Nano boards up to Macs and gaming GPUs
Worm across a test networkOn a firewalled network of 8 cloud nodes running Android, each with an unauthenticated debug (ADB) port, the open model GLM-5.2 replicated itself end to end, reaching all 8 nodes in 4 of 5 runs
Real IoT hardwareGoogle’s Gemma 4 26B model was split across 15 Android TV boxes with 4GB of memory each, running at about 0.39 tokens per second. Slow, but it worked
Botnet-scale mathsApplied to the Kimwolf botnet’s 2 million-plus devices, the researchers estimate roughly 133,000 model copies, equivalent to around 40 data-centre GPUs of inference
ForecastIf small models keep improving at their historical rate, they reach the “can replicate” level by the end of 2026

Two things struck me. First, the replicator doesn’t need root access. Inference can run entirely in user space, and the cheap devices it would target can simply be reinfected after a reboot. Second, it’s slow: the researchers talk about a doubling time of days, not minutes. That isn’t comforting. It means it would grind away quietly in the background, which is exactly how most IoT compromises behave already.

Why IoT is the first target

The researchers are blunt that home and small-business IoT would probably be first in line, because it’s always connected and has historically had minimal security. They chose Android TV boxes for a reason. The Kimwolf botnet, which Synthient researchers tracked to more than 2 million infected devices last winter, was built mostly from unofficial Android TV boxes. Many of them shipped with proxy software preinstalled and the Android Debug Bridge left open with no password. Brian Krebs also found Kimwolf infections inside corporate and government networks, which tells you where those cheap boxes end up.

That’s the uncomfortable link. Kimwolf proved criminals can already round up two million cheap devices. Anthropic has now shown what an AI model could do with them: not just send junk traffic, but think, adapt to unfamiliar hardware and keep going without anyone at the controls. As the researchers put it, unlike ordinary malware, such a model “could also figure out unfamiliar devices on its own”.

It also lands on top of something the UK has already been warned about. In April, the NCSC and international partners published an advisory on China-nexus covert networks of compromised devices. It describes state-linked attackers hiding behind botnets made mostly of compromised small office routers and IoT kit. Today, those devices are used as proxies. Anthropic’s work suggests the next step is using them as compute.

Why we think differently about security depending on the device and where it sits

This is the bit I’ve been chewing over since the Smart Buildings Show. We don’t apply one standard to connected kit. We apply about five, mostly by instinct. A battery LoRaWAN sensor gets waved through. A cellular router gets scrutinised. The Android signage player in reception, the NVR in the comms cupboard and the “smart” panel on the boardroom wall get ignored completely. Anthropic’s research is a good excuse to make that instinct explicit, because it adds a new question to the list: does this device have spare compute that someone else could borrow?

Device classCould it host an AI worm?Where the real risk sits
Battery sensors (LoRaWAN, NB-IoT, LTE-M)No. Microcontrollers with kilobytes of memory and no inbound routeData integrity, keys and where the readings land. LoRaWAN payloads are encrypted end to end, so the gateway can’t read them
LoRaWAN gatewaysUnlikely to host a model, but it’s a Linux box with internet backhaulTreat it as a router, because that’s what it is
Industrial cellular routersUnlikely. Most carry far less memory than a 4GB TV boxThey’re the path in and the proxy. Admin pages, default credentials, public IPs and unpatched firmware are what attackers want
NVRs, edge AI boxes, Jetson-class gatewaysYes. Real compute, often Linux, often forgottenAnthropic got inference running on Jetson Nano boards. If it runs AI for you, it could run AI for someone else
Android signage, TV boxes, smart panelsYes. This is exactly the hardware Anthropic usedOften bought on price, rarely patched, sometimes shipped with debug ports open. In smart buildings these are everywhere

That table also shows why the endless argument about where a device was made misses the point. In the replication scenario, what matters is function and exposure, not the flag on the box. A sensor from anywhere that can’t take inbound connections is a poor host. A well-known Western-branded device with an open debug port on a flat network is an excellent one. I’ll come back to the provenance question properly in a separate piece, because it deserves more than a paragraph.

Cellular has a quiet advantage, if you use it

Anthropic’s own blue-team section makes a point that cellular IoT people should take as a compliment. A worm like this would leave a network signature: multi-gigabyte model downloads to new hosts, plus sustained device-to-device traffic between machines that wouldn’t normally talk. In earlier research the authors cite, ISP quarantine of infected customers cleared 92% of infections within 14 days.

On a cellular estate, your operator or MVNO is your ISP, and you’ve got controls most broadband sites can only dream of:

  • Private APNs and no public IP. A device that can’t be reached from the internet can’t be recruited from the internet. I’ve made the case before that the public IP SIM was never the problem, but an exposed admin page on one is.
  • Data caps as an alarm. A SIM on a few megabytes a month suddenly pulling gigabytes is about as clear a red flag as you’ll ever get. Set the alerts and actually read them.
  • Traffic you can see. Managed connectivity platforms already show per-SIM usage and destinations. Sustained traffic between devices that should only ever talk to one server is worth a look.
  • Someone owning the router. Most of the exposure I see comes down to nobody being responsible for the kit after install. That’s the theme of who secures the router in the cabinet, and it’s never been more relevant.

What to do now

None of this needs new technology. It needs the basics done everywhere, including on the boxes nobody thinks of as IT:

  • Kill default passwords and open debug ports. Anthropic specifically points to the UK’s ban on default IoT passwords as the kind of rule that helps. PSTI covers consumer kit, but there’s no reason commercial estates should accept less.
  • Find your compute-capable devices. NVRs, signage players, Android panels, edge AI boxes and gateways. If it has gigabytes of memory and a network port, it belongs on the asset register and the patching schedule.
  • Segment. Building systems, cameras and signage shouldn’t share a flat network with each other, let alone with the office.
  • Watch for the signature. Large unexpected downloads, new device-to-device flows and outbound scanning. Your firewall, router or connectivity platform can probably already flag these.
  • Buy on more than price. The Kimwolf devices were cheap for a reason. A box with no update path and preinstalled proxy software is a liability from day one.

Keep it in proportion. Anthropic kept the hacking step deliberately trivial, with no authentication on the debug port, so its results are an upper bound. The end-of-2026 forecast is based on a general capability index, not a direct measure of replication. No AI worm has been seen in the wild. This is an early warning, not an incident.

My take

Some critics have called the research scaremongering, or good PR for an AI company. I don’t buy that as a reason to ignore it. Strip out the AI and the message is one this industry has heard for a decade: cheap, unmanaged, always-on devices are someone else’s infrastructure waiting to happen. Mirai proved it in 2016, Kimwolf proved it again last winter, and the NCSC told us in April that state actors are running whole networks of hijacked routers and IoT.

What Anthropic adds is a reason to stop judging devices only by the data they handle. A people counter that reports occupancy every fifteen minutes is low risk on data. An Android panel in the same building, with 4GB of memory and an open debug port, is high risk on compute, even though it handles no data at all. If your security policy only asks “what does it know?”, it’s now asking half the question.

Quick answers

What did Anthropic find about self-replicating AI?

Anthropic’s Frontier Red Team reported on 12 October 2026 that small AI models may be months away from being able to copy themselves across cheap IoT devices without a data centre. In simplified tests, the open model GLM-5.2 replicated itself across a test network, and Google’s Gemma 4 26B ran split across 15 Android TV boxes.

Is there a self-replicating AI worm in the wild?

No. Anthropic says it has not shown a full rogue replicator, its hacking step was deliberately simplified, and no such worm has been detected. The research is a capability warning, not a report of an attack.

Which IoT devices are most at risk?

Devices with real compute and poor security: cheap Android TV boxes, signage players, smart panels, NVRs and edge AI boxes, especially with default passwords or open debug ports. Battery sensors lack the memory to host a model, and well-managed cellular routers on private APNs are hard to reach.

How do you protect an IoT estate against AI worms?

The same way you protect it against botnets: remove default passwords, close debug ports, keep devices off the public internet or on private APNs, segment networks, patch firmware and monitor for unusual downloads and device-to-device traffic.

What is the Kimwolf botnet?

Kimwolf is an Android botnet that infected more than 2 million devices, mostly unofficial Android TV boxes, many shipped with proxy software preinstalled and an open Android Debug Bridge. Anthropic used its size to estimate how much AI compute a similar network could provide.

Sources: Anthropic Frontier Red Team, “AI models may be capable of self-replicating without access to a datacenter”, William Zhou and Daniel Freeman, 12 October 2026; Al Jazeera and Reuters coverage, 13 October 2026; NCSC and partners, “Defending against China-nexus covert networks of compromised devices”, April 2026; KrebsOnSecurity, “Kimwolf botnet lurking in corporate, govt networks”, January 2026; Synthient research on Kimwolf via The Hacker News and SecurityWeek, January 2026. Experiment figures are as reported by Anthropic and have not been independently replicated.