Confirmed *Anthropic Frontier Red Team blog Oct 12 2026* by William Zhou & Daniel Freeman *Times Now / Reuters / WAM Oct 13.*$GNS $WAL $TRUTH
Not eliminating need for data centers it's about *Autonomous Replication & Adaptation (ARA) worm risk* and not happening yet.
Finding:*Our experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial IoT devices, like smart ovens and off-brand Android TV boxes"_
- *What tested:*
*9B model* got inference running on *8 of 9 hardware types* tested
*Gemma4-26B-A4B (26B total params)* run across *15 cheap Android TV boxes ∼$55 each* = *0.39 tokens/sec* — slow but working — each box holds slice of layers — proves cheap hardware can host capable model by splitting
- Models *≤30B* can complete many individual steps: find vulnerable devices, install AI software, transfer files, distribute copies — but *struggle to chain full end-to-end* autonomously
- In *simplified evals (security steps removed, hacking trivial)*, *GLM-5.2 and stronger* completed end-to-end replication across 8 devices in 4/5 trials — every infection is worm hopping via *unauthenticated ADB port*, setting up inference server to serve *Llama3.1-8B*
- *Timeline:* Projection based on *13.7 ECI/yr trend, R² 0.95 proxy* — *small models could hit simplified replication level by end of 2026* if historical improvement continues — small weight-based replicators soon after — *no full rogue replicator built yet*
- *Why scary:* Unlike Mirai/Kimwolf botnet (Kimwolf = *2M+ devices, mostly off-brand Android TV boxes* template), AI worm can *improvise on unfamiliar hardware*, *passes on adaptations* to future copies (gets better), *no central switch-off*, doubling time *days* — Anthropic math: Kimwolf-size fleet = *∼133k replicas, ∼40 H100s worth inference* — devices: smart fridges, ovens, robot vacuums/lawnmowers.
Not eliminating need for data centers it's about *Autonomous Replication & Adaptation (ARA) worm risk* and not happening yet.
Finding:*Our experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial IoT devices, like smart ovens and off-brand Android TV boxes"_
- *What tested:*
*9B model* got inference running on *8 of 9 hardware types* tested
*Gemma4-26B-A4B (26B total params)* run across *15 cheap Android TV boxes ∼$55 each* = *0.39 tokens/sec* — slow but working — each box holds slice of layers — proves cheap hardware can host capable model by splitting
- Models *≤30B* can complete many individual steps: find vulnerable devices, install AI software, transfer files, distribute copies — but *struggle to chain full end-to-end* autonomously
- In *simplified evals (security steps removed, hacking trivial)*, *GLM-5.2 and stronger* completed end-to-end replication across 8 devices in 4/5 trials — every infection is worm hopping via *unauthenticated ADB port*, setting up inference server to serve *Llama3.1-8B*
- *Timeline:* Projection based on *13.7 ECI/yr trend, R² 0.95 proxy* — *small models could hit simplified replication level by end of 2026* if historical improvement continues — small weight-based replicators soon after — *no full rogue replicator built yet*
- *Why scary:* Unlike Mirai/Kimwolf botnet (Kimwolf = *2M+ devices, mostly off-brand Android TV boxes* template), AI worm can *improvise on unfamiliar hardware*, *passes on adaptations* to future copies (gets better), *no central switch-off*, doubling time *days* — Anthropic math: Kimwolf-size fleet = *∼133k replicas, ∼40 H100s worth inference* — devices: smart fridges, ovens, robot vacuums/lawnmowers.