Anthropic warns that smaller AI models could self-replicate across IoT devices, dynamically adapting exploits rather than running static code like traditional worms.
According to the company's research, these small models used programs that had access to hacking tools. The process included locating devices, gaining access, performing operations, and jumping to new targets.
The key difference compared to previous threats is the ability of the systems to learn from their mistakes. When an attempt failed, it left notes and instructions for subsequent attempts, so that the same misguided actions would not be repeated.
Setting up the operation on each new device proved to be the most difficult stage, as it requires many sequential decisions depending on the hardware and software of each target.
READ ALSO
Figma introduces new AI tools for faster design
However, none of the lightweight models tested were able to complete the entire propagation chain from start to finish. The process was also extremely slow, with a typical example being a test on 15 Android TV devices, where the speed reached just 0,39 data segments per second. This translates to a setup time of 71 to 150 hours for the unoptimized Gemma4-26B-A4B and 29 to 51 hours for Qwen3.5-9B. The researchers reported that the larger GLM-5.2 model appears to be able to perform the process reliably.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)