AI researchers warn of self-replicating malware risks across connected devices

Yesterday 21:21

By: Azzat Manal

AI researchers warn of self-replicating malware risks across connected devices

Researchers are investigating how artificial intelligence models could potentially spread across connected devices in ways that resemble computer worms. These malicious programs can replicate themselves across networks, adapt to different systems and pass information about previous attempts to subsequent copies, raising new questions about the security risks associated with increasingly capable AI technologies.

Experiments described by researchers associated with Anthropic’s Frontier Red Team explored how AI models could operate across networks of relatively inexpensive devices. In one test, a Gemma model was distributed across 15 Android TV boxes, with each device handling part of the workload and exchanging data with others. Although the process was slower than running a model on powerful servers, the experiment demonstrated that certain AI tasks can be distributed across modest hardware rather than relying entirely on high-performance computing infrastructure.

Other simulated experiments examined whether AI models could navigate networks, identify targets, transfer files and execute tasks across multiple nodes. In one test involving eight simulated Android nodes, the GLM model reportedly reached all nodes in four out of five attempts and seven nodes in the remaining attempt. Researchers also observed that passing information about previous failures between model instances could improve performance, with reported success rates for one model increasing from 45% to between 61% and 84%.

However, these findings should not be interpreted as proof that AI systems can currently infect devices across the public internet without human involvement. The experiments were conducted under specific conditions, including environments with limited or absent security protections, and did not demonstrate uncontrolled propagation across real-world networks. Their results instead illustrate potential capabilities that warrant further investigation.

The risks could become more significant as smaller AI models become increasingly efficient and can run directly on consumer electronics. Smart refrigerators, ovens, robot vacuum cleaners, lawnmowers, surveillance cameras and video-conferencing systems could represent potential targets if they contain vulnerable software, accessible network connections or sufficient computing resources to support AI functions.

Vehicles, drones and advanced robotic systems could also become relevant to future security assessments as onboard processors grow more powerful. Meanwhile, larger AI systems could potentially target data centres, where extensive computing resources and interconnected infrastructure could increase the consequences of a successful intrusion.

Researchers have warned that the capabilities demonstrated in controlled tests may develop further as AI models improve. Some projections suggest that smaller models could acquire more advanced capabilities by the end of 2026, although such forecasts remain uncertain and do not establish that autonomous, self-propagating attacks will become widespread.

Cybersecurity specialists continue to emphasise the importance of timely software updates, strong authentication, network segmentation and monitoring for unusual traffic. Manufacturers and operators of connected devices may also need to strengthen security testing as AI becomes more deeply integrated into everyday technology. The central concern is not that an uncontrolled AI worm has already emerged, but that increasingly capable models could make certain forms of automated cyberattack easier to coordinate if adequate safeguards are not in place.