Published October 15, 2026 · AI News
Anthropic says small AI models could be months away from copying themselves across cheap internet-connected devices, with no data centre needed. The warning comes from a research post by its Frontier Red Team, published October 12, 2026, and was reported by Reuters and SecurityWeek on October 13.
No rogue AI is spreading today. Anthropic says no full self-replicator exists yet. But its tests suggest the individual steps are getting easier for models small enough to run on a $55 TV box.
What changed
Anthropic bought 15 off-brand Android TV boxes (about US$55 each) and split a 26-billion-parameter open model across them. It ran, slowly: 0.39 tokens per second. In a closed test network of eight emulated Android devices, one open-weight model, GLM-5.2, completed the full copy-and-spread cycle in 4 of 5 runs.
| Finding | Detail |
|---|---|
| Models tested | Open-weight models of 30B parameters or fewer |
| Full replication | Only GLM-5.2 finished end to end (8/8 devices in 4 of 5 runs) |
| Hardware | 15 Android TV boxes at ~US$55 each; a 9B model ran on 8 of 9 device types |
| Speed | Slow: one replication takes 1–2 days for a 9B model |
| Timeline | Small models could reach GLM-5.2’s level by the end of 2026 if the trend holds |
Not everyone is alarmed. Ciaran Martin, former head of Britain’s National Cyber Security Centre, told Reuters the post had “more than a hint of vendor exaggeration.” A Georgetown researcher called it better read as a hypothetical than a prediction.
What it means for you
For people using ChatGPT, Claude or Gemini, nothing changes today. This is about open models that anyone can download, not the chatbots in your browser.
The practical takeaway is old advice that now matters more: cheap, no-name smart devices are the weak link. Anthropic’s own defence list starts with secure defaults, like no shared default passwords, and with internet providers spotting unusual traffic between home devices.
- Change default passwords on routers, cameras and TV boxes.
- Install firmware updates, or retire devices that no longer get them.
- Be wary of very cheap Android TV boxes from unknown brands.
For buyers in Canada and the U.S., a known brand that publishes security updates is the safer pick. The U.K. already bans default passwords on connected devices, a rule Anthropic points to.
Sources
- Anthropic: AI models may be capable of self-replicating without access to a datacenter
- Reuters: AI models may soon self-replicate without access to data centers, Anthropic warns
- SecurityWeek: Small AI models may soon be able to self-replicate on IoT devices
- Ars Technica: Anthropic warns small AI models could self-replicate through IoT devices