Anthropic’s research shows open-weight models setting up AI across cheap Android TV boxes, with a 30-billion-parameter threshold in view.
Anthropic’s AI self-replication warning says small models could copy themselves across smart devices by the end of 2026. The company published research on 13 October 2026. The research shows open-weight models setting up AI on cheap Android TV boxes. TF discusses the experiments, the limits and the defences.
What’s Happening & Why It Matters
The Warning
Anthropic says models with 30 billion parameters or fewer could replicate on consumer internet-of-things devices by the end of 2026. BleepingComputer reports the estimate assumes small models keep improving at their historical rate. Targets include smart ovens, TVs and TV boxes.
A replicator on such devices would have no data centre to switch off, Euronews reports. Anthropic says spread could be slow, with a doubling time of days. Defenders would have to find and clean each device.
The Experiments
Researchers networked 15 off-brand Android TV boxes, which Gizmodo prices at about $55 (€49) each with 4GB of RAM. The cluster ran Google’s open-weight Gemma4-26B-A4B model without an external data centre.
In a second test, Qwen3.5-9B set up Llama3.1-8B on 8 of 9 hardware types. GLM-5.2 reached all eight nodes of an emulated network in 4 of 5 trials. Qwen’s success rate rose from 45% to 84% when the model kept notes from earlier attempts. Opus 4.8 as an adviser gave no measurable gain.
The Limits
Anthropic did not show a self-running worm on real devices. The targets had open Android debugging access and lax security. Setup took 28 to 51 hours for Qwen3.5-9B and 71 to 150 hours for Gemma4-26B-A4B, which gives defenders time.
Marco Figueroa, who leads Mozilla’s bug bounty programme, accused Anthropic of overstating the findings, Gizmodo reports. Anthropic calls a rogue replicator theoretical for the time being.
The Defences
Anthropic urges detection of large model transfers and odd device-to-device traffic. Anthropic backs honeypots to inspect payloads, quarantine of infected devices and stronger default security on IoT products. More of these devices ship with small AI chips, so the target pool grows.
My take: the warning joins a run of lab disclosures. Microsoft’s Satya Nadella urged an AI emergency brake, as TF covered in Microsoft’s Nadella and Smith Call for an AI Emergency Brake. A replicator on TV boxes would escape any brake kept in a data centre.
TF Summary: What’s Next
Researchers will try to repeat the experiments on real devices. Device makers face pressure to close open debugging ports. Regulators will ask whether IoT rules cover models running on edge hardware. Anthropic has to show evidence for the end-of-2026 date.
MY FORECAST: Security firms will publish tests of AI worms on consumer devices by March 2027. The UK and EU will cite the research in IoT security reviews. Device makers will ship stricter default settings. Critics will keep disputing the timeline.
Related Stories
- Microsoft’s Nadella and Smith Call for an AI Emergency Brake
- TF Cybercrime Round-Up: 6-11 October 2026
- Meta Releases Muse Wonder, Its First Frontier AI Model