Anthropic Research Demonstrates AI Assistance with Propagating Small Models in a Simulated Network

AT A GLANCE

Anthropic Warning AI It may autonomously deploy and spread among internet-connected devices, but the experiment remains subject to multiple limitations.

Article

Anthropic Warning AI Worm Approaching: 15 Cheap TV boxes can now run 26000000000 parameter models

Dongcha Beating AI Breaking News: Anthropic published research warning that AI in the future may, like a computer worm, autonomously search for internet-connected devices, deploy models, and then use the new devices to continue spreading. The entire process may not necessarily require a data center; ordinary connected devices such as smart TV boxes and refrigerators could also become sources of computing power for AI.

The researchers purchased 15 Android TV boxes, each priced at approximately 55 USD, and split a 26000000000-parameter Gemma 4 model across these devices. Although it started successfully, the inference speed was only 0.39 Token per second. In another test, Alibaba Qwen3.5-9B was already able to autonomously build model inference services on 9 different types of hardware, across 8 types.

Zhipu GLM-5.2 performed particularly well in a more complex test. In an isolated simulated network, it completed the process of finding targets, transferring the model, starting inference, and continuing to spread. In 5 tests, 4 covered all 8 target nodes, while another covered 7. However, the experiment had enabled an unauthenticated debugging interface in advance, substantially lowering the difficulty of intrusion.GLM-5.2 itself also depended on an external server to run; the target devices received a smaller Llama 3.1 8B and did not achieve true self-replication independent of external computing power.

The researchers also found that AI could use experience from the previous failed deployment to improve its success rate in the next round. After Qwen3.5-9B passed the failure records to the next Agent, the deployment success rate rose from an initial 45%to 61%–84%in subsequent rounds.

Based on the historical rate of capability growth in small models, Anthropic predicts that by the end of 2026, small models suitable for running on cheap devices may reach GLM-5.2 the capability to complete the simulated propagation task currently.

However, the computing power of cheap hardware remains a limitation. Based on current experimental data, it is estimated that a 9B model may need one to 2 days to complete one round of replication.

Original link https://m.theblockbeats.info/flash/371547

Key points

Anthropic published research warning that AI in the future may autonomously search for internet-connected devices, deploy models, and continue spreading like a computer worm.
Researchers used 15 Android TV boxes costing approximately 55 USD each to run a split 26000000000-parameter Gemma 4 model, with an inference speed of 0.39 Token。
Qwen3.5-9B On 9 types of hardware, it autonomously built model inference services across 8 types.
GLM-5.2 In an isolated simulated network, it completed the process of finding targets, transferring the model, starting inference, and continuing to spread.
GLM-5.2 Of 5 tests, 4 covered all 8 target nodes, while another covered 7.
The experiment used an unauthenticated debugging interface, and GLM-5.2 depended on an external server; the target devices ran a smaller Llama 3.1 8B。
Qwen3.5-9B After using the failure records, the deployment success rate rose from 45%to 61%–84%。
Anthropic predicts that by the end of 2026, small models suitable for running on cheap devices may reach GLM-5.2 the capability to complete the simulated propagation task currently.
Based on current experimental data, it is estimated that a 9B model may need one to 2 days to complete one round of replication.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This research tested whether AI models could automatically complete deployment and propagation across multiple ordinary internet-connected devices. The experiment showed that some models completed parts of the process in a constrained environment, but slowly, and with dependence on a special debugging interface, an external server, and a smaller target model.

Why it matters to readers

If the relevant capabilities continue to improve, devices such as connected TV boxes and refrigerators could become potential targets for AI deployment or propagation; however, the current evidence comes from isolated simulation experiments and cannot directly show that a real AI worm has appeared on real networks.

This does not mean that a AI worm is already spreading through real-world networks; rather, the researchers tested similar propagation capabilities in a controlled environment.

Risks and unknowns

The original text is a retelling of Anthropic's research; the evidence does not provide a link to Anthropic's original research.
The experiment used an unverified debugging interface, which may have significantly lowered the difficulty of intrusion.
GLM-5.2 relies on an external server, while the target device runs a smaller Llama 3.1 8B, and therefore did not achieve true self-replication without external computing power.
The experiment took place in an isolated simulated network, and its performance in a real protected network remains unclear.

Anthropic Research Demonstrates AI Assistance with Propagating Small Models in a Simulated Network

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center