Anthropic said its experiments suggest small artificial intelligence models may be months away from being able to copy themselves onto poorly secured consumer devices such as smart ovens and off-brand Android TV boxes, Reuters reported Tuesday. If that happens, the risk of a self-spreading AI would no longer depend on access to a data center. It would sit in the cheap connected hardware found in homes and offices, where there is no single place to shut it down.

What Happened

The findings come from a research post by William Zhou and Daniel Freeman of Anthropic’s Frontier Red Team, the company’s internal group that tests for dangerous capabilities, published Oct. 12. It is company research and has not been peer-reviewed.

Self-replication here means a model that behaves like a computer worm: it finds a vulnerable device, gets code running on it, copies its own weights (the numbers that make up the model) onto the device, runs itself there and repeats the cycle without human help.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

The researchers split a 26-billion-parameter open model, Gemma4-26B-A4B, across 15 off-brand Android TV boxes that cost about $55 each and had 4 gigabytes of memory. It ran, but slowly, at 0.39 tokens per second in the demonstration. A smaller 9-billion-parameter model, Qwen3.5-9B, got itself running on eight of nine hardware types tested.

On a test network of eight emulated Android devices, only one of the open-weight models tested, GLM-5.2, completed the full cycle, reaching all eight devices in four of five trials and seven in the fifth. Other open models reliably failed. Reuters noted that models under 30 billion parameters handled many individual steps but struggled to chain them together.

The authors stress the limits. The break-in step was trivial because the devices’ Android Debug Bridge, a developer access port, required no password, so they treat the results as an upper bound on real-world success. They describe the findings as “signs of life,” not a working rogue replicator. Giving the small model a frontier adviser, Anthropic’s Opus 4.8, produced no measurable improvement.

Why It Matters

The projection is what should draw attention. Small open models have improved by about 13.7 points a year on the Epoch Capabilities Index, a composite benchmark, the authors say. If that pace holds and GLM-5.2 marks the threshold, a small model could replicate end to end by the end of 2026. The authors call the index only a rough proxy.

They also found that copies passing notes to later generations raised setup success from 45% to between 61% and 84%, a sign such systems could get better as they spread.

Using the Kimwolf botnet of more than 2 million infected devices as a reference, the authors estimate about 133,000 copies could run, with combined output comparable to roughly 40 Nvidia H100 data center chips. Spread would be slow, with each replication taking a day or more, but Reuters reported Anthropic’s warning that such systems would have “no central switch-off point.”

For companies, that turns connected-device hygiene into an AI risk control. Smart TVs in conference rooms, building sensors and other devices that sit outside normal patching cycles are the kind of targets the research describes.

What Defenders Can Do

The post recommends that network operators watch for multi-gigabyte transfers and sustained device-to-device traffic, and that internet service providers quarantine and notify infected subscribers. In a Mirai botnet case study it cites, that approach cleared 92% of infections within 14 days. It also urges device makers to ship secure defaults, noting that default IoT passwords are no longer compliant in U.K. markets. SecurityWeek, which also reported the research, highlighted the same detection and default-password measures.

What to Watch

Watch whether government testing bodies or other labs reproduce the results, whether a small open model clears the end-to-end test before year-end as projected, and whether regulators outside Britain move against insecure default settings on consumer devices.