# Anthropic Tests Two Steps Toward an AI Worm, Not Real-World Infection

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/31926
Published: 2026-10-14T15:40:33.000Z
Updated: 2026-10-14T15:40:33.000Z
Section: Technology

> Tests ran a model across 15 Android TV boxes and simulated replication among cloud-hosted Android devices. They did not demonstrate a worm infecting real-world devices.

Anthropic’s Frontier Red Team tested whether AI models could run across low-cost devices and coordinate replication, demonstrating two steps that could contribute to a self-replicating computer worm but not a worm infecting real devices.

The team ran Gemma4-26B-A4B across 15 Android TV boxes priced at about $55 each. The distributed setup processed about 0.39 tokens per second.

In a separate experiment, researchers used eight Google Cloud Platform nodes running an Android 15 emulator. An unauthenticated Android Debug Bridge port simplified access, making the setup an upper bound on real-world success. Among the open-weight models tested, GLM-5.2 reached all eight nodes in four of five trials and seven of eight in the remaining trial. The replication test used Llama 3.1-8B as its payload.

The experiments tested distinct parts of a possible attack. The TV-box setup demonstrated distributed inference, while the cloud experiment tested replication in a simulated environment. The work largely set aside how a worm might first gain access to devices.

Anthropic estimated that setting up inference on its TV-box configuration could take 71 to 150 hours with Gemma4-26B-A4B and 28 to 51 hours with Qwen3.5-9B. Its projection that small models could reach the tested replication capability by the end of 2026 is an extrapolation from past capability trends, not a demonstrated result.

Researchers William Zhou and Daniel Freeman wrote, “We don’t yet have a full rogue replicator.”

May experiments showed AI agents exploiting vulnerable web applications, obtaining credentials and deploying an inference server with copies of their harness and prompt on compromised hosts. Those tests, like Anthropic’s, did not establish that an AI worm is spreading in the wild.