IT之家 · 智能时代 · 10/13/2026, 15:15:32
Anthropic Warns: Small AI Models May Self-Replicate on IoT Devices Within Months
Anthropic's experiments reveal that small AI models under 30 billion parameters can already perform multiple steps of self-replication, with end-to-end autonomous spread on commercial IoT devices expected within months. The study warns these agents bypass security checks, interpret unfamiliar devices, and target data centers, posing a shutdown challenge due to the lack of a kill switch.
SOURCE COVERAGEFull Original Report
October 13 — Anthropic stated on Monday that its experiments indicate small AI models could soon acquire the ability to self-replicate on commercial Internet of Things (IoT) devices, such as smart ovens and generic TV set-top boxes, within a few months.
In these experiments, Anthropic investigated whether AI models could identify vulnerable devices, copy themselves onto those devices, and continue spreading without human intervention.
The findings revealed that smaller models with fewer than 30 billion parameters can perform several steps in the self-replication process but struggle to independently execute the entire end-to-end workflow.
However, in simplified test environments where certain security checks were removed, Anthropic discovered that GLM-5.2 and more powerful models were capable of achieving full end-to-end self-replication.
Anthropic warned that such self-replicating models are difficult to shut down, as there is no single kill switch. Unlike traditional non-AI malware, these models can autonomously understand unfamiliar devices and even target data center processors for infiltration.
The study focused primarily on connected IoT devices such as Android TVs and smart home appliances, due to their generally weak security protections.
IT Home noted that the AI giant also cautioned that if small models continue to iterate at their historical pace, autonomous replication could be achieved by the end of this year. Following closely behind, smaller replicators based on model weights would also emerge.
Anthropic indicated that the replication speed of such autonomous AI systems will likely be slow and will rely on vast amounts of widely available, poorly secured computing resources. The report recommended strengthening system defenses to increase the difficulty of propagation for these AI agents.