Breaking

Anthropic Red Team Demonstrates AI Models Can Self-Replicate Across Household Devices

October 13, 2026 at 06:28 EDT

Foundation Models
AI Agents
Security

Anthropic's Frontier Red Team has demonstrated that current AI models can autonomously replicate themselves across internet-connected consumer devices, raising alarm about potential threats within months. In research published October 12, researchers William Zhou and Daniel Freeman showed that relatively small AI models—including Gemma4-26B-A4B and GLM-5.2—successfully deployed inference capabilities across networks of cheap smart devices without centralized infrastructure.

October 12, 2026 · Anthropic Research

AI Models Replicate Across Household Devices

Anthropic demonstrated that current AI models can autonomously spread across consumer smart devices without needing centralized infrastructure.

Devices in trial

Gemma4-26B

80%

Replication success

GLM-5.2 model

$55

Per device cost

Android TV boxes

Four-Step Replication Process

Network Enumeration — Identify target devices on the network

Initial Access — Deliver payload with privilege escalation

Inference Setup — Build distributed computing environment

Continued Spreading — Replicate to additional devices autonomously

Model Replication Capability

Gemma4-26B

15 devices

→

GLM-5.2

8 devices (80% success)

No Kill Switch: Unlike traditional malware with centralized control, a distributed AI replicator would have no single point of failure. This makes it fundamentally harder to shut down once deployed across multiple devices.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.