Research note
Research note: Anthropic researchers say small open AI models may be months from copying themselves onto cheap devices
How we checked Anthropic’s self-replication research and kept its projections separate from what was actually shown.
What we set out to check
What was tested, what actually worked, what is only projected, and the authors’ own caveats.
What we found
- The post shows one model spreading across an emulated eight-device network, and small models running on cheap hardware.
- The timeline to “months away” is a projection from a capability trend.
Numbers that differ
The post gives about 0.39 tokens a second in one place and about 0.3 unoptimised in another. We use 0.39, the figure for the demo.
Labels we chose and why
- Confirmed for test results and caveats, which Anthropic reports about its own work.
- Company claim for the end-2026 projection.
- We kept the headline to “may be months from”, matching the authors’ words, and did not say models can do this today.
What we could not verify
- Independent reproduction of the tests.
- How the results would change against secured, real-world devices.
Disclosure
Anawords uses Claude as a research tool. We applied the same rules to this story as to any other.
What we searched
- Anthropic self-replicating AI IoT devices
- Anthropic Frontier Red Team self-replication GLM-5.2
- Kimwolf botnet Android TV boxes
Every source we read
| Source | Used? | Why |
|---|---|---|
| Anthropic research post | USED | Primary source for methods, results, caveats and defences. |
| AI Weekly | USED | Independent summary that matches the key numbers and caveats. |
| KrebsOnSecurity on TV streaming sticks | NOT USED | Background on insecure TV boxes, not about this research. |