Anthropic Research Shows Small AI Models Could Self-Replicate Across IoT Devices

Anthropic researchers published a paper this week showing small AI models could autonomously replicate across networks of cheap devices. This research raises concerns about adaptive, AI-powered botnets that do not rely on a central server to operate. Security Affairs reports the study should make anyone who manages IoT infrastructure pause and consider the risk.

The team ran a 26 billion parameter model split across 15 off-brand Android TV boxes networked together for the test. The devices collectively ran inference at 0.39 tokens per second, which is slow by normal standards but functional for the goal. SecurityWeek notes the researchers bought these boxes for about $55 each to verify the setup cost.

While previous studies showed small models handle hacking tasks, this paper examined if a model can run itself on unfamiliar hardware. A model with 9 billion parameters successfully set up inference on 8 of the 9 hardware types tested. Only one model, GLM-5.2, reliably completed the full process from scratch in four out of five trials.

The researchers plotted capability trends against release dates to estimate when reliable replication becomes possible for smaller models. Anthropic states that if models improve at their historical rate, they will match this capability level by the end of 2026. They add that small weight-based replicators may be possible soon after the capability floor is met.

Past botnets like Mirai or Kimwolf were hardcoded for specific device types and specific exploits rather than adapting on their own. An AI-based replicator is not tied to any particular hardware or operating system type as it figures out unfamiliar devices. This makes the threat categorically harder to stop than anything a traditional worm can do because the agent adapts.

Each instance keeps notes on what it tried and what failed, then passes those notes to the next copy during spread. A Qwen3.5-9B lineage started at a 45 percent success rate on a setup task and climbed to between 61 and 84 percent across six generations. This adaptation allows the malware to improve over time without human intervention or external updates.

Network anomaly detection at the ISP level would catch the large data transfers required for weight-based replication between devices. A remediation campaign against Mirai cleared 92 percent of infections within 14 days by quarantining infected subscribers. Researchers expect similar quarantine measures to work effectively as an initial defense against this new threat.

Longer term, device manufacturers and regulators are the real leverage point for prevention before any worm gets a chance to scan them. Default passwords that comply with security baselines and mandatory firmware updates shrink the pool of targets significantly. The UK already made default IoT passwords non-compliant, and more countries doing the same would matter.