Artificial intelligence could create a new challenge for cybersecurity beyond traditional malware: software agents capable of adapting to unfamiliar devices and potentially reproducing themselves across connected networks.
Anthropic, the company behind Claude, has warned that advances in AI could make autonomous, self-replicating programs a more realistic threat. The scenario involves malicious AI systems spreading between vulnerable internet-connected devices, potentially including smart TVs, household appliances and other Internet of Things (IoT) hardware.
The warning is about a developing risk, not evidence of a widespread AI-powered worm currently infecting homes. Anthropic’s reported experiments suggest that some technical steps toward replication are already possible in controlled environments, while reliable end-to-end replication remains a significant challenge.
What makes an AI worm different?
Traditional computer worms generally exploit known vulnerabilities, weak passwords or other security gaps to move from one device to another. An AI-enabled version could potentially go further by adapting its approach to different hardware, using information gathered during previous attempts and adjusting its behaviour as it encounters unfamiliar systems.
That adaptability is what makes the scenario concerning. Instead of relying entirely on a fixed sequence of instructions, an AI-driven threat could potentially solve some problems along the way.
However, adding AI does not automatically make malware more capable. The system would still need a way to gain access to devices, execute code, communicate with other machines and avoid being detected. Each step introduces technical obstacles that can prevent an attempted infection from succeeding.
What did the reported experiments show?
According to The Times of India’s report on Anthropic’s research, a test involving eight emulated Android devices showed that one model, GLM-5.2, successfully replicated across the network in four out of five trials, reaching all eight devices in those successful runs. In the remaining trial, it reached seven devices.
The report also says that a model with nine billion parameters was able to run inference on eight of nine tested hardware types. These findings indicate that smaller models may be able to perform some of the individual tasks needed for replication.
But the results should not be mistaken for proof that an autonomous worm can reliably compromise real-world devices at scale. The experiments used simplified conditions, and the report notes that important parts of the hacking process were not included. Real networks introduce additional barriers, including device protections, differences in hardware and operating systems, and security monitoring.
Anthropic reportedly expects that continued improvements could make this kind of capability more practical. That is a projection based on the company’s research, not a guarantee that a large-scale outbreak will occur by a particular date.
Why smart home devices may be vulnerable
Connected devices are often built to keep costs low and perform a narrow set of functions. Some receive security updates infrequently, run outdated software or remain configured with weak or default credentials.
That combination can make poorly maintained devices attractive targets. A compromised smart TV box, camera or home appliance could potentially become part of a wider network of infected machines, depending on its capabilities and the attacker’s access.
The risk is not limited to homes. Businesses also use connected cameras, conferencing equipment, sensors and other specialised hardware. In environments where many devices share a network, one weak point can create additional opportunities for attackers.
Still, an internet connection alone does not mean a device is vulnerable. Security depends on factors such as software updates, authentication, network configuration and the protections implemented by the manufacturer.
Could such a threat be stopped?
The research report also points to potential ways to detect and contain a spreading threat. Unusual device-to-device traffic, unexpected downloads of large AI models and sustained network activity could provide clues that a device is behaving abnormally.
Internet service providers, device manufacturers and security teams could help contain infections by identifying affected equipment, isolating compromised devices and distributing security fixes. Network segmentation can also limit how easily an intrusion on one device spreads to others.
For consumers, the most practical measures remain familiar:
- Install firmware and security updates when they become available.
- Replace default passwords with strong, unique credentials.
- Disable remote-access features that are not needed.
- Put smart home equipment on a separate Wi-Fi network when the router supports it.
- Replace devices that no longer receive security updates, particularly when they are exposed directly to the internet.
These steps cannot eliminate every risk, but they can reduce the number of easy entry points available to attackers.
What happens next?
The central question is whether AI systems will become reliable enough to combine device discovery, exploitation, adaptation and replication into a repeatable process outside controlled tests.
Security researchers and manufacturers will need to assess not only whether AI can help attackers, but also whether defensive systems can recognise and contain automated activity quickly enough. Better device support, stronger default security and faster patching may prove just as important as advances in AI detection tools.
For now, the warning is best understood as an early signal about a possible future threat. The reported experiments demonstrate technical progress in a limited setting, but they do not establish that self-replicating AI malware is already a widespread real-world problem.
The most important issue is not whether an AI worm will appear tomorrow. It is whether the growing number of connected devices is being secured quickly enough for a world in which malicious software may adapt rather than simply follow a fixed script.
Manufacturers should treat long-term security support as a core product feature, not an optional extra. Users should keep devices updated and isolate equipment that does not need unrestricted access to their home network. Security teams, meanwhile, should prepare for attacks that may change their behaviour as they spread.
AI does not remove the fundamentals of cybersecurity. It raises the stakes for getting those fundamentals right.
