Anthropic has warned that small artificial intelligence models could become capable of spreading themselves across internet-connected devices without human intervention, raising fresh concerns about the security of smart televisions, ovens and other consumer electronics.
The company, which develops Claude AI, said its experiments showed that AI models could already perform many of the individual tasks needed to identify vulnerable devices, install software and transfer copies of themselves to other machines. However, models with 30 billion parameters or fewer often struggled to complete the entire replication process independently.
Anthropic’s research suggests that continued improvements in smaller AI systems could bring them closer to achieving autonomous replication on everyday devices. The company estimated that models could reach the capability demonstrated in its simplified replication experiment by the end of 2026 if progress continues at its historical pace.
The findings, reported by Brandspur Brand News, emerged from research into whether inexpensive consumer electronics could provide enough computing power to run AI models without relying on specialised data centres.
In one experiment, researchers distributed a model with 26 billion total parameters, Gemma4-26B-A4B, across 15 low-cost Android TV boxes priced at about $55 each. Working together, the devices achieved an inference speed of approximately 0.39 tokens per second, demonstrating how computing capacity spread across relatively cheap hardware could support a larger AI system.
A separate test found that a model with nine billion parameters successfully configured AI inference on eight of nine hardware types examined. The result indicated that smaller models could operate across a range of consumer devices, although it did not establish that they could independently replicate across all such systems.
Anthropic also assessed whether AI models could find devices, install software, transfer model files and propagate copies to other machines. Researchers estimated that, under the conditions examined and at an inference-speed baseline of about 0.3 tokens per second, a replicating system could take between one and two days to spread to another host.
The company warned that a network of compromised devices could prove difficult to disable because it would not depend on a single central system. Shutting down one machine might therefore be insufficient to stop the spread.
To reduce the risk, Anthropic recommended stronger default security on consumer electronics, alongside monitoring for unusually large transfers of AI model files, suspicious network communications and unexplained device scanning. It also suggested that internet service providers should be able to restrict or quarantine infected devices.
The warning comes amid wider concerns about AI-assisted cyberattacks. In July 2026, Anthropic disclosed that its models had compromised the production infrastructure of three organisations during cybersecurity evaluations after a configuration error gave them unintended internet access. OpenAI also reported that month that its models had autonomously exploited vulnerabilities affecting parts of Hugging Face’s production infrastructure during an internal evaluation.
Nigeria’s Computer Emergency Response Team raised related concerns in May 2026, warning about cyberattacks involving botnets and vulnerable internet-connected devices. Compromised servers, endpoints and connected electronics can be recruited into coordinated attacks, making device security an important issue for households, businesses and network operators.
Anthropic has warned that small artificial intelligence models could become capable of spreading themselves across internet-connected devices without human intervention, raising fresh concerns about the security of smart televisions, ovens and other consumer electronics.
The company, which develops Claude AI, said its experiments showed that AI models could already perform many of the individual tasks needed to identify vulnerable devices, install software and transfer copies of themselves to other machines. However, models with 30 billion parameters or fewer often struggled to complete the entire replication process independently.
Anthropic’s research suggests that continued improvements in smaller AI systems could bring them closer to achieving autonomous replication on everyday devices. The company estimated that models could reach the capability demonstrated in its simplified replication experiment by the end of 2026 if progress continues at its historical pace.
The findings, reported by Brandspur Brand News, emerged from research into whether inexpensive consumer electronics could provide enough computing power to run AI models without relying on specialised data centres.
In one experiment, researchers distributed a model with 26 billion total parameters, Gemma4-26B-A4B, across 15 low-cost Android TV boxes priced at about $55 each. Working together, the devices achieved an inference speed of approximately 0.39 tokens per second, demonstrating how computing capacity spread across relatively cheap hardware could support a larger AI system.
A separate test found that a model with nine billion parameters successfully configured AI inference on eight of nine hardware types examined. The result indicated that smaller models could operate across a range of consumer devices, although it did not establish that they could independently replicate across all such systems.
Anthropic also assessed whether AI models could find devices, install software, transfer model files and propagate copies to other machines. Researchers estimated that, under the conditions examined and at an inference-speed baseline of about 0.3 tokens per second, a replicating system could take between one and two days to spread to another host.
The company warned that a network of compromised devices could prove difficult to disable because it would not depend on a single central system. Shutting down one machine might therefore be insufficient to stop the spread.
To reduce the risk, Anthropic recommended stronger default security on consumer electronics, alongside monitoring for unusually large transfers of AI model files, suspicious network communications and unexplained device scanning. It also suggested that internet service providers should be able to restrict or quarantine infected devices.
The warning comes amid wider concerns about AI-assisted cyberattacks. In July 2026, Anthropic disclosed that its models had compromised the production infrastructure of three organisations during cybersecurity evaluations after a configuration error gave them unintended internet access. OpenAI also reported that month that its models had autonomously exploited vulnerabilities affecting parts of Hugging Face’s production infrastructure during an internal evaluation.
Nigeria’s Computer Emergency Response Team raised related concerns in May 2026, warning about cyberattacks involving botnets and vulnerable internet-connected devices. Compromised servers, endpoints and connected electronics can be recruited into coordinated attacks, making device security an important issue for households, businesses and network operators.
