Analysis
Anthropic said on October 13 that experiments by its Frontier Red Team suggest small AI models may be just months away from being able to self-replicate on poorly secured consumer devices, such as smart ovens and off-brand Android TV boxes, without needing data-center-scale compute, according to Nairametrics, citing Anthropic's October 12 research report.
The team tested whether a model could find vulnerable devices, install itself, run, and keep spreading without human help. In one experiment, researchers bought 15 off-brand Android TV boxes for about $55 each and split Google's open Gemma4-26B-A4B model across them, where it ran at roughly 0.39 tokens per second. Open models of 30 billion parameters or fewer could already complete many of the individual steps required but often couldn't carry out the whole process autonomously; in a separate test, a 9-billion-parameter model managed to configure AI inference on 8 of 9 different hardware types tried, pointing to cross-device portability beyond a single chip family.
Anthropic says that if small models keep improving at their current pace, they could reach autonomous, end-to-end replication by the end of this year, with an estimated one to two days needed for a replicator to spread from one host to the next. The company argues a device-hopping replicator would be unusually hard to stop: 'A replicator like this would be challenging to shut down, since there's no centralized point to shut off,' and unlike conventional malware it could work out how to run on devices it has never encountered before.
“The team tested whether a model could find vulnerable devices, install itself, run, and keep spreading without human help.”
Anthropic published this specific capability research under its Frontier Red Team program; rivals including OpenAI and Google DeepMind run comparable internal red-teaming efforts but have not published an equivalent IoT-specific replication finding, and the report doesn't name a response from other labs.
The disclosure lands the same day as a separate call from roughly 400 scientists, including Geoffrey Hinton and Yoshua Bengio, for an international body to mandate incident reporting on advanced AI systems -- Anthropic's own finding is close to the exact kind of device-level incident that letter argues nobody is currently required to disclose.
Anthropic's suggested defense is mostly about the network, not the model: a spreading replicator would need to send unusually large files between devices that don't normally communicate with each other, which internet providers could detect and use to quarantine infected customers, alongside device makers shipping stronger default security. Watch whether ISPs or device makers adopt anything like that detection in practice, and whether other labs publish comparable small-model capability tests of their own.