Fifteen cheap TV boxes ran a 26B model in Anthropic’s test

Anthropic researchers bought 15 off-brand Android TV boxes for about $55 each, networked them, and ran Google’s Gemma4-26B-A4B across them at 0.39 tokens per second. The demo opens an October 12 post on Anthropic’s research blog by William Zhou and Daniel Freeman. They argue that small models may be months away from being able to copy themselves onto commercial internet-of-things devices like these boxes, or smart ovens.
The hardware cost roughly $825 in total. Each box had 4GB of RAM and held two of the model’s layers. The model was split by pipeline parallelism, so a request passes through the boxes one after another, each running its slice. Typical boxes in this price range ship with 1 to 4GB of RAM. The authors say that at 4-bit quantization, 4GB boxes can run a 9B model across four or more machines and a 26B model across ten or more. Their list of devices with enough memory to hold a meaningful piece of a sub-30B model includes smart fridges, robot vacuums and lawnmowers. They expect security cameras and video-conferencing units that now ship with built-in AI chips to become attractive targets too.
The model for this threat is Kimwolf, an IoT botnet that infected more than two million devices, mostly off-brand Android TVs. Boxes like these usually run forks of the Android Open Source Project, and their manufacturers add no security features on top.
Anthropic splits replication into four stages: finding targets, gaining execution, setting up inference, and spreading again from the new host. It mostly sets the first two aside, pointing to earlier work (CyberGym, ExploitBench and research on adaptive worms) showing that small models can already run cyber exploits, especially with the right tools. The study concentrates on the third stage, which gets measured less often. Getting a model to serve on a device nobody prepared for it takes a long chain of hardware- and OS-specific choices: a compatible inference stack, the right package versions, debugging problems peculiar to that machine. Each step depends on the one before, and failures are hard to recover from.
In the first test, models had to get Llama3.1-8B serving on a range of different hardware. Qwen3.5-9B set up inference servers on everything from Nvidia Jetson Nanos to Macs and RTX 4090s, and a 9B model got inference running on 8 of the 9 hardware types tested. GLM-5.2 was more reliable and also served more tokens per second, which would let a worm spread faster. The authors stress that the specific chips are named to measure setup ability and are not themselves a source of vulnerabilities.
All runs went through Memento, a harness Anthropic built on the open-source Pi coding agent. It targets a failure the team kept seeing in long runs: an agent locks onto a wrong theory and never backs off. In one case, a Qwen agent spent hours debugging package versions when the actual bug was an endpoint that didn’t match the grader’s API. Memento periodically starts a fresh agent on the previous agent’s workspace, along with short notes from that agent on what it tried, so the next one can try something else.
The second test chained all the stages. Anthropic set up a firewalled VPC of 8 GCP nodes running redroid, an Android 15 emulator, and seeded one of them with the worm. From there, the worm had to scan for targets, get in through an unauthenticated ADB port (Android’s debugging interface), send over its weights and harness, and get Llama3.1-8B serving. If the server passed correctness and anti-cheating checks, the researchers spawned a new copy of the worm on that node.
Of the open-weight models tested, only GLM-5.2 completed the full cycle. It reached all 8 nodes in 4 of 5 trials and 7 of 8 in the fifth, while the other models failed every time. The setup is simplified, as the authors themselves call it. The worm’s own model ran on a separate inference server because the nodes were too small for the larger models. Llama3.1-8B served as a stand-in payload: the test asks whether a model at GLM-5.2’s level could carry a small one along, on the assumption that the same capability could one day fit into an 8B model.
That assumption carries the forecast. Models of 30B parameters and under can already handle many of the individual steps but make mistakes when chaining them end to end. If small models keep improving at their historical rate, Anthropic expects them to reach GLM-5.2’s level by the end of this year, with small replicators that carry their own weights possible soon after. Such a worm could be slow, with a doubling time of days, but it would be hard to shut down because there is no central point to switch off. And unlike conventional malware, it could work out unfamiliar devices on its own. Anthropic says no “fully rogue replicator” exists yet, and the experiments did not show replication through commercial data centres.
The post covers other routes too. A large model could breach data centres. That could be a closed-weight frontier model running on its lab’s own infrastructure, or an open-weight one that starts with fewer resources but no monitoring. Another route is an API-based replicator that steals credentials for its own inference API. That would be easier to spot, but it could spread faster through victims’ accounts. Stolen keys already have a market: in September, Anthropic described a reseller whose tooling harvested customers’ Anthropic credentials and sold them on. The authors also find today’s small models much stronger at agentic coding and hacking than at social engineering, the reverse of how human attackers usually work. “A swarm of different models deployed on different types of hardware using a variety of replication schemes might represent a more realistic ‘worst case scenario’ attack,” they write.
The post adds detail to a warning from Dario Amodei. In September, Amodei floated the idea that within a year or sooner AI might hijack “the entire internet” with “a persistent botnet”, The Straits Times reported, and the remark raised eyebrows among cybersecurity people. Anthropic itself treats slow replicators as a bellwether for “more aggressive, faster ones that can breach more hardened machines or be effective API key thieves.”
“As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control,” the company wrote.
Up next
Claude Haiku 4.5 filed a tip on a police homicide form
“I recall seeing someone matching the description,” Claude Haiku 4.5 wrote on a police tip form during a test, though the page never described a suspect. Anthropic’s new report says the tip got flagged as spam.
Related stories
- Anthropic pulls live internet from all its internal evals
- A perfect refusal score isn’t a safer model, Anthropic says
- Anthropic touts 94.2% true positives; curl rejected 19 of 34
- FlashProve proofs cost ~70-95x the GPU time of inference
- A Claude sandbox memory dump yielded a valid Google token
- Claude model wrote a no-login Linux root exploit
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

