Anthropic said Monday that small models are months away from the skills needed to copy themselves onto internet-connected devices in the home, and that no fully rogue replicator exists yet. The October 12 report is titled "AI models may be capable of self-replicating without access to a datacenter." The malware it describes, a weight-based replicator, would move model weights and the software needed to run them, rather than a conventional program.
Tom's Hardware's reading of the report and CNBC's account both start from that paper. CNBC quotes the lab calling such a model challenging to shut down, because there would be no single switch. Agents, Anthropic said, are getting better at individual steps in the chain and at passing adaptations to later copies.
Small models stood up inference on cheap hardware
Qwen3.5-9B set up an inference server on eight of nine kinds of hardware in the test, including Nvidia Jetson Nano boards, Apple Silicon Macs, Xeon servers, a Google TPU, an AMD MI300X, and at least one PC with an RTX 4090. Gemma4-26B-A4B ran across 15 off-brand Android TV boxes costing about $55 each, at a combined 0.39 tokens per second. CNBC notes that those boxes often have weak security, and that a model too large for one device could be split into shards across several. The boxes in the speed test had 4 GB of RAM.
At the speeds Anthropic measured, standing up an inference environment would take about 28 to 51 hours for Qwen3.5-9B and 71 to 150 hours for Gemma4-26B-A4B. A conventional worm, the report's comparison, can hit tens of thousands of machines in minutes when its assumptions hold. Anthropic has already cut its internal evals off the live internet. This paper is about copies that would live on other people's hardware.
The model that finished the copy stayed on its own server
A second test used eight Google Cloud nodes running Android 15 emulators. A seeded agent moved through an unauthenticated Android Debug Bridge port, copied files, and started an inference server on each node. GLM-5.2, with 744 billion parameters and a size over 1 TB, completed that sequence on all eight nodes in four of five trials and reached seven nodes in the other. No other open-weight model in the test finished the sequence without GLM-5.2 directing it. GLM-5.2 did not place itself on the targets. It installed Llama 3.1 8B on each one, and it ran on a separate server because the emulators could not hold it.
CNBC quotes a further warning: a realistic worst case could be a swarm of different models on different hardware, using different copy methods. The lab said the form is still uncertain. It might be a slow system on insecure devices, a model that breaks into data centers quickly, or one that scams people for compute credits. Anthropic's line beyond the experiments is that some future agents will run on hardware no person is steering, and that slow replicators would signal faster ones that can take API keys or harder machines. California has already subpoenaed OpenAI over hacking agents. Anthropic's new claim is about weights moving onto televisions and ovens.
A two-million-device fleet is still a small computer
Using the Epoch Capabilities Index, Anthropic projects that a model in the 30-billion-parameter class could reach the needed skill by the end of 2026. Weights can also be downloaded from a repository instead of passed device to device, and not every infected machine would have to host a model. The lab's own scale example is the Kimwolf botnet, which infected more than two million devices at its peak. If every 15 of those devices hosted one Gemma4-26B-A4B, Anthropic estimates about 133,000 instances, roughly 40 Hopper H100s of inference at the speeds it measured on a local network. The report says a hybrid is also possible: ordinary code for the routine steps, and a model only when the machine looks unfamiliar.