AI could spread through cheap TV boxes
In short
- Fifteen inexpensive Android TV boxes collectively ran an AI model.
- AI can combine technical steps to run on unfamiliar hardware.
- The experiments did not prove unrestricted spread across the internet.
- Cyberattacks can cost businesses millions and indirectly affect investors.
Fifteen inexpensive TV boxes were able to run an AI model together. Anthropic demonstrated this in research published on October 12, 2026. The researchers also examined how AI could independently perform the steps needed to run on other devices and continue spreading. This does not mean that an uncontrollable AI worm already exists. It does show how cyberattacks could take a different form in the future.
Fifteen inexpensive TV boxes ran an AI model
Anthropic used fifteen Android TV boxes costing around 55 dollars each. The total hardware cost was approximately 825 dollars, excluding additional expenses.
Together, the devices ran Gemma4 26B A4B at approximately 0.39 tokens per second. A token is a small piece of text processed by an AI model. That is slow for a chatbot, but speed was not the objective. The experiment tested whether inexpensive devices could collectively perform AI computations without a large data center.
The researchers also tested different hardware types. A model with 9 billion parameters successfully ran AI inference on 8 of the 9 hardware types tested.
This is a technical demonstration, not proof that every device is vulnerable. It shows that AI does not have to depend exclusively on large servers.
AI can combine several stages of an attack
A traditional computer worm is malicious software that copies itself to other computers. An AI agent could add another capability. It might attempt to understand unfamiliar devices and solve technical problems.
Anthropic examined four stages: finding devices, gaining access, running the model and continuing to spread.
According to the research, models with up to 30 billion parameters could already complete many individual steps. Combining every stage remained more difficult. In simplified tests, GLM 5.2 and stronger models could complete the full process.
The researchers stressed that it remains difficult to prove that such a system can spread reliably through a real network. The initial stages of an attack were not the main focus of their experiment.
That distinction matters. The results do not prove that AI can easily bypass security controls. They show that AI can combine technical operations that previously required more manual work.
North Korea, Iran and China could benefit
Countries that conduct cyber operations could benefit from technology that automates parts of an attack. North Korea has been associated with financially motivated cybercrime. Iran and China have also been subjects of research into state linked cyber activities.
An AI system that investigates devices and attempts to adapt to different circumstances could theoretically support espionage, data theft or disruption of digital services.
Consider poorly secured cameras, routers and televisions. If compromised devices then run AI computations, they could form a distributed network that does not depend on one central server. This could make it harder to shut down every component at once.
The research does not establish that North Korea, Iran or China is already using this specific method operationally. These are possible applications, not confirmed attacks.
Criminal groups could also take an interest. An infected device might quietly become part of a network used for further cyberattacks.
A cyber incident can cost millions
Cyberattacks create costs through downtime, lost revenue, recovery work and legal obligations.
IBM reported average global data breach costs of 4.44 million dollars in 2025. The average in 2024 was 4.88 million dollars, a decline of 9 percent. The figures come from research by IBM and the Ponemon Institute. IBM sells cybersecurity services and has a commercial interest in this subject. These figures are averages, not forecasts of the cost of an AI attack.
A concrete example comes from Lee Enterprises, a US media company. In a filing with the US Securities and Exchange Commission, the company reported that a cyber incident in February 2025 caused 10.5 million dollars in cash flow losses.
Investors need to distinguish between damage to a company and the market reaction. A business may lose revenue, incur recovery costs and lose customers. The stock market then determines how heavily those consequences affect the company's valuation.
Suppose you own 10,000 euros worth of shares. If the share price falls by 15 percent, your position is worth 8,500 euros. Your unrealised loss is 1,500 euros.
If the price falls by 30 percent, the same investment is worth 7,000 euros. Your loss is 3,000 euros. These are mathematical examples, not predictions of how markets will react to cyberattacks.
Europe requires manufacturers to report security problems
The European Union introduced the Cyber Resilience Act to improve the security of products with digital components. The law sets requirements for manufacturers that design, maintain and support digital products.
Since September 11, 2026, reporting obligations have applied to certain actively exploited vulnerabilities and serious security incidents. In applicable cases, manufacturers must submit an initial warning within 24 hours and a full notification within 72 hours.
The main remaining obligations apply from December 11, 2027. The exact rules depend on the product and the relevant requirement.
Regulation alone is not enough. Older devices may remain connected to the internet for years after manufacturers stop providing updates. Default passwords and unnecessary open connections also increase risk.
The European Commission has also published practical guidance for manufacturers and developers. The aim is to make digital security part of product development rather than something addressed only after an incident.
Organisations must secure ordinary devices too
Anthropic's experiments do not prove that an AI worm can already spread without restriction across the internet. The tests took place in controlled conditions, and important parts of a real attack were simplified.
Nevertheless, cameras, routers and televisions deserve attention. They can become entry points when security is inadequate.
Organisations can reduce risks by updating devices, changing default passwords and blocking unnecessary network access. Monitoring unusual traffic and quickly isolating compromised devices can also help.
Investors should understand how dependent a company is on digital systems. Prolonged downtime can affect revenue, costs and reputation. This does not automatically predict a share price, but it makes cybersecurity relevant when assessing business risks.
Questions to ask
- Which devices no longer receive security updates?
- Can an organisation quickly isolate compromised devices?
- How much revenue could be lost if systems remain unavailable for several days?
- How does a company assess the security of suppliers and cloud services?
- What financial consequences could a cyber incident have for a company and its shareholders?
This article is for educational purposes only and does not constitute financial advice.
