Anthropic researchers utilized the Petals technology, originally developed by Yandex, to investigate AI security risks. The project involved collaboration between Yandex, HSE University, the University of Washington, Hugging Face, and École Normale Supérieure Paris-Saclay. Petals enables the execution of large language models by distributing computation across multiple devices connected via the internet, bypassing the need for expensive dedicated servers.
In a specific experiment, Anthropic scientists used distributed inference to run the 26-billion-parameter Gemma 4 model across 15 Android TV boxes. This test confirmed that large neural networks can function on low-power hardware without a centralized server. The researchers also explored whether AI agents could autonomously propagate between devices, conducting tests in an isolated network with intentionally weakened security.
The findings aim to assess the risks associated with autonomous AI systems and develop countermeasures. Beyond Anthropic's study, Petals has been adopted in other international research initiatives. In 2026, researchers from the University of California, Merced, applied the technology to improve distributed inference efficiency, while scholars from Tsinghua University referenced Petals in their work on decentralized AI infrastructure.
