Anthropic researchers say small AI models may be months from self-replicating on cheap devices
In tests, a 9B model got inference running on 8 of 9 hardware types, and a model spread across a simplified 8-node network in most trials. The authors stress they don't yet have a full rogue replicator.
Anthropic published research on October 12, by William Zhou and Daniel Freeman, arguing that "small AI models may be months away from being capable enough to self-replicate." The work measures how close today's small open models are to copying and running themselves on inexpensive devices without help.
What the tests showed
"In our evaluations, a 9B model got inference running on 8 of the 9 hardware types we tested." In a simplified eight-node network, GLM-5.2 "spread to 8/8 nodes in 4 out of 5 trials and 7/8 in 1." In a cost demonstration, the team ran Gemma4-26B-A4B "distributed over 15 inexpensive off-brand Android TV boxes (~$55 each)," generating "0.39 tokens per second."
The limits the authors put on it
The paper is explicit that this is not yet the full threat: "We don't yet have a full rogue replicator." The spread test used devices with no authentication, "which makes this an upper bound on real-world success."
Why it matters
For most companies the near-term takeaway is inventory, not alarm: small models now run on hardware cheap enough that nobody budgets or tracks it. Knowing which models run where in your organization, including on devices outside the AI line item, is the same visibility that keeps AI spend accountable.