Anthropic has warned that AI models could soon copy themselves across internet-connected devices without human help. Smaller AI models could spread through vulnerable devices and create new cybersecurity risks.
In a report published on October 12, 2026, Anthropic’s Frontier Red Team warned about the risks posed by small AI models. The team said these models could learn to replicate themselves across everyday connected devices within months.
These include smart ovens, Android TV boxes and other Internet of Things (IoT) devices.
Such systems could behave like computer worms, spreading across networks and using infected devices to run AI models. They could also continue copying themselves without human involvement.
Unlike traditional malware, AI-based threats could adapt to different devices and find ways to run on unfamiliar hardware. This could make them harder to detect and stop.
However, Anthropic has not shown that an AI system can replicate itself across the public internet without human help. Its research shows that AI models are developing some of the skills needed for self-replication. However, major technical challenges still remain.
Anthropic Tests Show AI Models Can Complete Key Self-Replication Tasks
Anthropic’s research found that AI models with 30 billion parameters or fewer can already carry out several tasks needed to copy themselves to other devices. These tasks include finding target devices, transferring model files and installing the software required to run the model on another machine.
However, models still struggle to complete the entire process reliably. In Anthropic’s tests, the open-weight model GLM-5.2 spread across all eight test nodes in four out of five trials. In the remaining trial, it reached seven of the eight nodes.
The researchers tested the model on a controlled network of eight cloud-based nodes running an Android emulator. The setup included an unsecured debugging port, which made it easier for the model to move between devices.
How AI Self-Replicating Models Could Spread Across Devices
Traditional computer worms spread between devices without human help. Anthropic warns that AI-based worms could do the same by finding vulnerable devices, gaining access and copying themselves to other machines.
AI models could also adapt to different hardware, potentially making them harder to stop. In one experiment, Anthropic ran a 26-billion-parameter model across 15 low-cost Android TV boxes, each with 4GB of RAM. Although the system was slow, the test showed that multiple inexpensive devices could work together to run a large AI model.
ALSO READ: ARTEX AI Agent Pulled From GitHub After Cyberattacks Target South Korean Banks
Why AI Self-Replicating Worms Could Be Hard to Stop
Anthropic warns that self-replicating AI could be difficult to shut down if it spreads across multiple devices. Removing the model from one machine may not stop copies running elsewhere.
Such systems could target poorly secured smart TVs, security cameras, household appliances and other internet-connected devices. Smaller AI models may pose a greater concern because they need fewer computing resources to run. They could also split their workload across several devices when one machine lacks enough memory.
However, Anthropic noted that an AI worm might spread slowly. Instead of infecting thousands of devices within minutes, it could take days to double the number of infected machines.
Anthropic Warns Small AI Models Could Self-Replicate by the End of 2026
Anthropic estimates that small AI models could gain the ability to copy themselves across devices without human help by the end of 2026 if current progress continues.
However, this is a prediction, not a confirmed deadline. Even if AI models learn how to replicate themselves, they would still need access to vulnerable devices to spread.
This has highlighted the need to improve cybersecurity as AI systems become more capable. Anthropic urged security teams to address weak passwords, outdated software and poorly protected connected devices to reduce the risk of AI-based attacks.
ALSO READ: AI Cyber Attacks Statistics [2025 and beyond]
Anthropic’s AI Self-Replication Warning Draws Mixed Reactions
Anthropic’s warning has drawn mixed reactions from cybersecurity experts. While some see AI self-replication as a growing threat. Others question whether the company’s tests support its predictions.
Reuters reported that Ciaran Martin, former head of the UK’s National Cyber Security Centre, questioned whether Anthropic’s demonstration justified the scale of its warning.
Jessica Ji is a senior research analyst at Georgetown University’s Center for Security and Emerging Technology. She said the possibility of AI systems replicating themselves deserves attention, but it remains unclear how likely this is to happen.
The debate shows that an AI model succeeding in a controlled test does not mean it can carry out large-scale cyberattacks in the real world. Anthropic’s data point to a possible future security risk. They do not show that self-replicating AI systems are already spreading across the internet.
