.
Artificial intelligence company Anthropic has warned that smaller AI models could become capable of replicating themselves across internet-connected household devices within months, raising fresh cybersecurity concerns about the potential spread of autonomous AI systems beyond traditional data centres.
In a research report published on October 12, 2026, the company said its experiments indicate that compact AI models may soon be capable of identifying vulnerable devices, installing AI software on them and continuing to spread without direct human intervention.
The researchers cautioned that if smaller models continue improving at their historical rate, they could reach the capability demonstrated in a simplified replication experiment by the end of 2026.
The findings raise concerns about the security of internet-connected devices, including smart televisions, ovens, refrigerators and other consumer electronics that may have limited computing resources but inadequate security protections.
Smart TVs and Household Appliances Could Become Targets
Anthropic’s research examined whether inexpensive consumer devices could provide sufficient computing power to run AI models without relying on specialised data-centre infrastructure.
The company tested its approach using 15 low-cost Android TV boxes, each costing approximately $55 and equipped with 4GB of RAM. Researchers distributed the Gemma4-26B-A4B model across the devices, allowing them to work together to run AI inference.
The experiment demonstrated that relatively inexpensive hardware could collectively support a model that would otherwise require more substantial computing resources.
Anthropic said devices such as smart ovens, refrigerators, automatic vacuum cleaners and lawnmowers could potentially host portions of AI models, particularly when several machines are connected and their computing resources are coordinated.
The research also found that a nine-billion-parameter model could run inference across eight of nine hardware types tested, highlighting the growing ability of smaller AI systems to operate on unfamiliar devices.
Researchers Demonstrate Key Steps Towards AI Replication
The study assessed four stages involved in autonomous replication: finding potential targets, gaining access to a device, establishing an environment in which an AI model can run, and continuing the replication process from the newly accessed machine.
Anthropic said models with 30 billion parameters or fewer can already complete many of the individual tasks required for replication, although combining those steps into a reliable, uninterrupted process remains challenging.
In the company’s simplified evaluations, GLM-5.2 and stronger models demonstrated end-to-end replication capabilities under the experimental conditions.
However, Anthropic distinguished these results from proof that a fully autonomous AI worm can reliably spread across commercial devices in real-world environments.
The researchers also noted that their investigation concentrated heavily on the ability to establish AI inference on unfamiliar hardware. They did not attempt to comprehensively demonstrate every cybersecurity exploit that would be necessary to compromise arbitrary devices.
The company’s projection that smaller models could achieve the required capability by the end of the year is based on historical improvements in model performance. It is therefore an estimate, rather than a confirmed deadline for the emergence of such systems.
Why AI Worms Could Be Different From Conventional Malware
Traditional botnets generally target particular operating systems, devices or known vulnerabilities. Anthropic warned that AI-powered malware could be more adaptable because models may be able to work out how to operate on unfamiliar hardware and adjust their approach as circumstances change.
Such systems could also distribute model components across multiple devices instead of depending on a single, powerful computer.
The absence of a central control point could make a successful AI replicator more difficult to contain. Even if individual devices were disconnected or cleaned, other compromised machines might continue operating.
Anthropic nevertheless estimated that replication across low-powered consumer hardware could be slow. Depending on the model and configuration, its experiments suggested that some replication-related tasks could take between one and several days.
The threat, therefore, may be less about an instantaneous takeover of connected devices and more about persistent, distributed activity that gradually expands across vulnerable networks.
Cybersecurity Risks Extend Beyond Data Centres
The findings highlight a potential weakness in the expanding Internet of Things (IoT) ecosystem, where televisions, home appliances, cameras and other devices increasingly connect to the internet.
Many inexpensive devices have limited memory, receive inconsistent security updates or lack the safeguards commonly found in enterprise computing systems.
Anthropic pointed to the Kimwolf botnet, which reportedly infected more than two million devices, primarily Android television boxes, as an example of how poorly secured consumer hardware can be assembled into a large network.
The researchers used that scale as a reference point for estimating how an AI-enabled network might operate. Their calculations suggested that a fleet of comparable devices could support many distributed AI model instances, although the resulting computing capacity would remain modest compared with modern data-centre infrastructure.
The company stressed that these estimates were illustrative and based on a specific class of hardware and a previous botnet example.
Network Monitoring and Device Security Could Limit the Threat
Anthropic said defenders could reduce the risk by preventing a replicating system from spreading to enough additional devices to sustain itself.
One important measure would be monitoring network traffic for unusual communication patterns associated with devices attempting to locate, access or transfer software to other machines.
Manufacturers and network operators could also strengthen device security through timely software updates, restricted access, improved authentication and tighter controls over the installation of unauthorised software.
The broader objective is to reduce the number of new devices compromised by each infected machine until the replication process can no longer sustain itself.
AI Safety Debate Enters a New Phase
Anthropic’s findings add to growing concerns about how increasingly capable AI systems could be misused in cybersecurity attacks.
Although the research does not establish that autonomous AI worms are already spreading through household devices, it suggests that falling hardware requirements and improving model capabilities could make decentralised AI replication more feasible.
For technology companies, device manufacturers and internet service providers, the warning reinforces the need to treat security as a central part of AI deployment and connected-device design.
The critical question is whether cybersecurity safeguards can advance quickly enough to match improvements in AI capabilities. If smaller models become better at adapting to unfamiliar systems and reproducing themselves, the security of everyday internet-connected devices could become an increasingly important part of the global AI safety debate.











