Anthropic ran AI on 15 Android TV boxes at $55 each. What to check if you own a similar box

Fifteen Android TV boxes, each costing about 55 USD and holding just 4 GB of RAM, were linked together to run one AI model of roughly 26 billion parameters. The speed was about 0.39 tokens per second. A token is the chunk of text a model reads and writes, so each chunk took two to three seconds. That is a test on real hardware by Anthropic’s Frontier Red Team (the group that checks risks at the edge of model capability), published on 12 October 2026.

The test sits inside a larger warning from Anthropic: small AI models “may be only months” away from being good enough to copy themselves onto commercial IoT devices, meaning internet-connected things such as TV boxes, cameras or appliances. Read closely, the part worth a household’s attention is not the chatbot on a phone. It is the cheap box plugged in behind the TV that nobody remembers updating.

What was measured, and what is an estimate

Anthropic split the experiment into parts, and the confidence differs between them:

• Measured on real hardware. A 9-billion-parameter model set up a model-serving machine on its own on 8 of the 9 hardware types tested. Splitting a Gemma4 model of about 26 billion parameters across 15 Android TV boxes was also a real run. Anthropic’s page gives 0.39 tokens per second in the introduction and about 0.3 in the appendix.
• Simulated. The “self-spreading” part ran in an 8-node network built from Android emulators in the cloud, with one ADB port (Android’s debugging bridge) left open without authentication. Only the GLM-5.2 model managed it reliably: it spread to 8/8 nodes in 4 of 5 runs and to 7/8 nodes in the remaining run. The other models failed consistently. Anthropic says the attack step here was so easy that the result is a ceiling. It also points out that earlier IoT botnets such as Mirai spread using nothing more than lists of manufacturers’ default usernames and passwords, so real-world IoT security may not be much harder than the easy step in the test.
• Extrapolation. The “months” timeline is inferred from the trend in model capability, and Anthropic itself calls it only a rough estimate. The number of copies at botnet scale is also a calculation: if a swarm of infected TVs the size of the Kimwolf botnet at its peak (over 2 million devices, mostly Android TV boxes) were grouped 15 boxes at a time, each running one 26-billion-parameter Gemma4, that gives about 133,000 copies and, at most, the compute of roughly 40 H100 GPUs. Anthropic calls this a rough estimate too, based on one type of hardware and one old botnet.

Al Jazeera, reporting on 13 October, notes that in the tests no model was yet both strong enough and small enough to copy itself reliably onto a smart TV. Professor Roman Yampolskiy (University of Louisville) is quoted as saying this is not an existential threat today, but that the trend is worrying.

Who Anthropic says should do what

Anthropic’s defence list is mostly aimed at network providers and manufacturers: watch for unusual traffic, such as multi-gigabyte model transfers or devices calling each other to run inference, isolate infected devices and notify customers, and require manufacturers to ship with safe default settings. On notification, it cites experience with the Mirai malware: when customers were notified, 92% of infections were cleaned up within 14 days; when they were not, 74%. On default settings, it notes that the UK has already treated default passwords on IoT devices as non-compliant.

So what can an ordinary user do? Their share is small but real: do not be the open door. The steps below rest on Google and Android documentation. We have not run each step on a specific TV box, and menu names differ from brand to brand.

Three questions for every Android box at home

1. Which company made it, and when was its last system update? Google says a device that is not Play Protect certified “may not be secure”, may not receive system or app updates, and its Google apps are not licensed versions. Google’s page is written for phones: tap the profile picture in the Play Store, then go to Settings and About to see the certification status. It does not describe how to do this on a TV. If you cannot say which company made the box, or the box has no update entry, that answer is already information.
2. Is debugging turned on? According to Android documentation, the Developer options screen is hidden by default, and you reveal it by tapping Build number seven times. USB debugging and Wireless debugging both live inside it. If you see that screen showing on a box someone else set up, and nobody remembers why, ask whoever installed it, or look at it and turn both debugging options off. The Android documentation is written for phones, and menu names on a TV box differ by brand. The ADB port Anthropic simulated is a network port, and the documents we checked do not say how to close it for each brand, so this article does not promise one step for every box.
3. Do you still need this box? A box that no longer gets updates and is used only to watch films could be replaced by a device from a clearly identified manufacturer, or unplugged from the network when it is not in use. That is the author’s judgement, not a recommendation from Anthropic.

What to take away

Anthropic’s warning is a conditional forecast, based on a simulated test with an easily exploited weakness. Nothing in this article shows that the TV box in your home has been, or is about to be, taken over. But the test ran on exactly this kind of 55 USD box, and the three questions above take a few minutes. If you can answer all three for each box, you have done your part. If a box fails to answer, it is the one to watch before small AI gets good enough.

Sources

1. Anthropic — research on self-replicating AI: the primary source from 12 October 2026: the 15-box Android TV test, the 8-node simulation results, the limits Anthropic states itself and its defence list.
2. Al Jazeera — AI could self-replicate on home devices within months, Anthropic says: the 13 October 2026 report, noting that no model reliably copied itself onto a smart TV in the tests, with the view of an independent researcher.
3. Google — Check whether your Android device is Play Protect certified: explains that uncertified devices may not be secure and may not get updates; the steps are written for phones.
4. Android Developers — Configure on-device developer options: says Developer options is hidden by default and that USB debugging and Wireless debugging sit inside it.