Anthropic: small AI models may be months from self-replicating without a datacenter
What's going on? Anthropic's Frontier Red Team argues that small AI models may be months away from being capable enough to self-replicate on commercial IoT devices such as smart ovens and off-brand Android TV boxes — without needing a datacenter. Such a weight-based replicator would be hard to shut down centrally, though Anthropic says it may be slow (doubling time measured in days).
What they measured: In simplified evaluations, models at or under about 30B parameters already complete many individual replication steps; GLM-5.2 was the only open-weight model in their set that reliably finished end-to-end replication across a firewalled fleet (spreading to 8/8 nodes in 4 of 5 trials). A 9B model got inference running on 8 of 9 hardware types tested. If small models keep improving at their recent rate, Anthropic's trend line puts a small model at GLM-5.2's replication capability level by the end of 2026 — a rough estimate, they stress, because the Epoch Capabilities Index is only a proxy.
Why it matters: The company frames this as early "signs of life" for AI worms: coding-agent skills generalise to cyber tasks, including autonomous replication. It also shows a 26B model running distributed inference across 15 ~$55 Android TV boxes. Defences it highlights include ISP anomaly detection, quarantine playbooks used against Mirai-style botnets, and harder IoT defaults.
The pushback: David Sacks argued on X on 14 October that the experiments don't support the dramatic headlines. He pointed out that the 26B model ran across the 15 TV boxes at about 0.39 tokens per second, that the end-to-end test used emulated devices with deliberately open ADB ports (making the "hacking" step trivial), that the capable model often ran on a separate server, and that agents were harnessed and told to propagate. He also noted that AI-free worms already do this at scale, from Mirai in 2016 to Kimwolf, which has infected more than two million devices, mostly cheap Android TV boxes with open ADB ports. Replit CEO Amjad Masad went further in a reply, calling the result "fake": he said the hacking was done by GLM, and that the model was swapped for a small Llama for the self-replicating stage because GLM was too big.
Read more: Anthropic, "AI models may be capable of self-replicating without access to a datacenter" (William Zhou & Daniel Freeman, 12 Oct 2026) · David Sacks on X · Amjad Masad on X
