RESEARCH · AI AGENTS · CYBERSECURITY
Can AI copy itself? What Anthropic’s experiment showed
Anthropic, the company behind the Claude chatbot, is studying whether AI could move itself to other devices and keep running without a person. In one experiment, the team connected 15 ordinary TV boxes and ran a model across them. Here is why they did it, what they tested and why this does not mean AI is already spreading through people’s televisions.
Why use 15 TV boxes for AI?
A TV box is a small computer connected to a television to play videos and run apps. The researchers wanted to see whether combining several low-powered devices would provide enough resources to run AI.
Gemma4-26B-A4B ran across 15 boxes at 0.39 tokens/second. A token is a short piece of text. That is slow for a conversation, but it showed that this hardware could run the model.
This does not mean each box hosted its own complete chatbot. In this experiment, one model was split across the devices, which worked together.
Anthropic: original research and methodology, 12 October 2026 ↗
What does “AI copies itself” mean?
Imagine moving an application from your laptop to another computer and launching it there. You now have two running copies. The software has not become smarter; it simply has another place to run.
It helps to distinguish a model from an agent. A model generates responses. An agent is software that uses a model and tools to take actions, such as reading files or running commands. Copying a file is not enough for independent propagation: the system also needs access to another device and a way to run there.
This is not evidence of consciousness or a desire to “survive.” Launching another copy and developing intentions are different things. Automated software with excessive permissions can cause harm without either.
Did the AI copy itself independently?
Propagation was tested separately in an isolated network with simplified access. An AI assistant set up the Llama3.1-8B model on another device. The researchers’ test software launched the next assistant copy; the model directing its actions ran separately.
The experiments demonstrated important parts of the process, not a fully independent “AI worm” taking over arbitrary home devices without the researchers’ help. The authors also say they do not yet have a complete rogue replicator.
Anthropic: original research and methodology, 12 October 2026 ↗
Malware already spread on its own. What is new?
A computer worm moves between devices without someone manually copying it. Worms existed long before ChatGPT. A well-known example is Mirai, a botnet that exploited weak protection on connected devices. Researchers documented it in a USENIX study.
AI could add the ability to adapt to unfamiliar environments: examine an error and change the next action. A separate laboratory paper, AI Agents Enable Adaptive Computer Worms, studies this approach across different systems. It is another study, not a report that your network has been infected.
Here is a simple way to think about it: a fixed set of instructions may stop at an unfamiliar error. An agent may try to understand it and find another way forward. That does not give it universal access to every computer.
USENIX: Understanding the Mirai Botnet ↗
AI Agents Enable Adaptive Computer Worms: a separate paper, 2 June 2026 ↗
What should an ordinary user do?
You do not need to throw away your television. This article provides no evidence that your devices are infected. A useful step is to check whether your TV box, router and cameras receive official updates, whether default passwords have been changed and whether unnecessary remote administration is enabled.
NCSC guidance covers updates and protection for smart devices. NIST’s catalogue helps assess concrete features: access controls, secure updates and security settings. If a manufacturer has stopped providing support, take that into account when you next replace a device.
If you are building an AI assistant, the lesson is more specific: grant access only to what the task requires. An assistant fixing a game does not need every cloud account, billing setting and device in your home. OWASP calls unnecessary tools and permissions an Excessive Agency risk.
Our practical example: start with the project files and a test environment. Before a long task, set a spending limit, permitted changes and a stop condition. These boundaries help, although they do not guarantee complete security.
NCSC: using smart devices safely at home ↗
What could happen next?
Our conclusion: this is an early signal for developers and device manufacturers. Weak protection on home devices was already a problem; autonomous software could make it harder to manage. The issue is who can run software on a device and what that software is allowed to do.
Our forecast, not an experimental result: if small models become better at long tasks, permission controls and visibility into their actual actions will matter more. This research does not establish a date for a widespread “AI worm.”
The practical takeaway is to keep devices updated, limit unnecessary access and distinguish a laboratory experiment from an established widespread infection.
Quick answers
Does self-copying mean AI has become conscious?
No. Software propagation and consciousness are different questions. Launching another copy is not evidence of consciousness.
Is every local model dangerous?
No. Assess the whole system: tools, permissions, network access and deployment controls. The label “local model” alone does not determine the risk.
Should I turn off all my smart devices?
This article is not an infection diagnosis. Start with official updates, passwords, manufacturer support and network settings; isolate a device and seek help if there is a specific concern.
Discussion
Share a thought or ask about the article. Your nickname and comment will be public.
Sign in with email
Get a one-time code. No password needed.
The code expires in 10 minutes. Signing in does not subscribe you to the newsletter. Data use.
Or sign in with ChatGPT ↗Loading comments…