Warning about a new class of cyber threats
Anthropic, the developer of the Claude AI model, has published the results of a study warning of the potential emergence of so-called "AI worms". According to expert estimates, these could be autonomous programs capable of spreading independently between various devices without human intervention. Potential targets include smart televisions, refrigerators, robot vacuums, and other household appliances.
Experiments with small models
Researchers tested whether compact neural networks could run on accessible hardware and independently configure their operation on different platforms. In one of the tests, the Gemma 4 model was launched on 15 inexpensive Android set-top boxes, each costing approximately 55 US dollars. In another experiment, the Qwen3.5-9B model successfully configured its launch on eight out of nine types of tested equipment.
The more powerful GLM-5.2 model demonstrated a more pronounced ability to spread: it organised the transfer of another neural network to all eight devices in four out of five trials. However, specialists note that the experiments were conducted in a simplified environment with limited protective mechanisms, and small models currently cannot reliably perform the entire process independently.
Self-improvement mechanism
An important aspect of the research was the discovery of how transferring records of predecessors' actions to new AI copies affects efficiency. This practice increased task completion success rates from 45% to 61–84%. This indicates the potential ability of such systems to refine their actions during the process of spreading.
Forecasts and differences from viruses
According to Anthropic's forecast, small models may reach a similar level of capabilities by the end of 2026. At the same time, researchers emphasise that fully autonomous self-replicating AI does not currently exist.
Unlike ordinary computer viruses, such systems could potentially adapt to unfamiliar equipment and spread without a single centre of control. However, the process may be quite slow: creating a new copy could take several days.
Anthropic warns that countering such threats will require joint efforts from device manufacturers and internet service providers.
Comments
No comments yet. Be the first!