New experiments suggest that smaller AI models are approaching the ability to replicate across networks of connected devices. Researchers say the risk deserves attention, although a fully autonomous AI worm has not yet been demonstrated.
By GAWAH News Desk | Technology
Artificial intelligence could introduce a new dimension to cybersecurity threats, with researchers warning that smaller AI models may eventually be capable of spreading themselves across internet-connected devices, including smart TVs and other household electronics.
In a research paper published on October 12, AI company Anthropic said its experiments indicate that relatively small AI models can already perform many of the individual tasks needed for autonomous replication. The company warned that continued improvements could bring these capabilities closer to practical use.
The findings raise questions about the security of the growing number of devices connected to the internet, particularly products that receive limited software updates or ship with inadequate security protections.
However, Anthropic has not demonstrated a fully autonomous AI worm spreading uncontrollably across the public internet. Its findings combine experiments conducted on physical devices with tests in a controlled, simulated network.
How could an AI worm work?
A computer worm is malicious software that can spread from one machine to another without requiring a person to install it on every device.
An AI-enabled version could potentially use an AI model to identify unfamiliar hardware, work out how to run software on it and adapt its approach when it encounters technical obstacles.
According to Anthropic’s research, the process involves several stages: identifying potential targets, gaining access, configuring an AI model to run on compromised hardware and continuing the replication process.
The researchers examined whether relatively small models could complete these tasks in sequence rather than simply perform them individually. Their results suggest that the underlying capabilities are advancing, although important limitations remain.
Why smart TVs and household devices matter
One striking aspect of the research is its focus on inexpensive internet-connected devices rather than powerful computers in large data centres.
Anthropic demonstrated that a model with 26 billion parameters could be distributed across 15 inexpensive Android TV boxes, with different devices handling portions of the computation. The experiment showed that a network of relatively modest devices could collectively run an AI model, albeit slowly.
In a separate evaluation using a simplified, firewalled test network, the researchers found that the open-weight model GLM-5.2 could complete the tested replication process across the network under the experimental conditions.
These tests do not establish that ordinary smart TVs are currently being infected by self-replicating AI. They demonstrate technical capabilities under controlled conditions, including assumptions that would not necessarily hold on a properly secured real-world network.
Nevertheless, the findings highlight why poorly secured connected devices could become attractive targets. An attacker might seek to exploit their combined computing resources, use compromised devices to attack other systems or create a distributed network that is harder to dismantle.
What makes AI-driven malware different?
Conventional malware can already spread automatically, and large botnets have previously compromised millions of internet-connected devices. AI-driven malware is therefore not an entirely new category of cyberattack.
The potential difference lies in adaptability. An AI-enabled threat could, in principle, help automate technical tasks that would otherwise require significant human effort. It might attempt to configure software for unfamiliar hardware, troubleshoot errors or adapt its methods to different environments.
If such capabilities became sufficiently reliable and could be combined with effective methods of gaining unauthorised access, the range of devices that attackers could target might expand.
Anthropic cautioned that smaller models can complete many individual steps today but still make mistakes when combining them into a complete replication process. The company projected that continued progress could close some of these gaps, potentially by the end of 2026.
That timeframe is a projection based on observed capability improvements, not a confirmed deadline for the emergence of an operational threat.
Researchers urge caution over the warning
The findings have also prompted questions about how quickly the threat might materialise. Reuters reported on October 13 that cybersecurity experts were divided over Anthropic’s assessment. Some considered the research a reason to strengthen defences, while others questioned whether the company’s warning overstated the immediacy of the danger.
Ciaran Martin, the former head of Britain’s National Cyber Security Centre, was among those expressing scepticism about the scale of the scenario described.
Anthropic itself acknowledged a central limitation: it has not yet demonstrated a complete, fully autonomous rogue replicator operating in the real world.
The distinction matters. Showing that an AI model can complete individual technical tasks, or spread through a simplified test network, is not the same as proving that it can independently compromise large numbers of securely configured devices across the public internet.
The research should therefore be understood as an assessment of an emerging risk rather than evidence of a widespread AI-worm outbreak.
What can users do to protect connected devices?
The research reinforces several established cybersecurity practices that households and businesses can adopt:
- Install software updates: Keep smart TVs, routers, cameras and other connected devices updated with security fixes.
- Replace default passwords: Use unique, strong passwords for device administration and change factory-default credentials.
- Secure the home router: Change its administrator password, keep its firmware updated and disable remote administration when it is not needed.
- Review connected devices: Disconnect products that are no longer supported or cannot receive essential security updates.
- Use a separate network where possible: Put smart-home devices on a guest or isolated network if the router supports it, limiting their access to personal computers and other sensitive equipment.
- Choose manufacturers carefully: Prefer products with a clear security-update policy and a history of maintaining their software.
These measures do not guarantee protection against every attack, but they can reduce exposure to common vulnerabilities and limit the damage caused by a compromised device.
The bigger cybersecurity question
As AI models become more capable of operating software and handling complex technical tasks, cybersecurity researchers will increasingly need to evaluate not only what these systems can do under supervision, but also what could happen if their capabilities were misused or combined with existing vulnerabilities.
Anthropic’s research offers an early look at that challenge. It demonstrates progress in the individual capabilities that could support autonomous replication, while leaving significant questions about reliability, real-world access and the speed at which such a threat might spread.
For device manufacturers, network operators and users, the immediate lesson is practical: improve the security of connected devices before attackers find new ways to exploit them.