Anthropic Warns AI Worms Could Spread Across Smart Devices
Summary
Anthropic says small AI models may eventually replicate across vulnerable IoT devices, but its tests did not demonstrate autonomous spread on compromised household devices. Researchers estimate the capability could emerge by the end of 2026.
Key points
- Anthropic ran a Google Gemma model across 15 off-brand Android TV boxes and outlined a four-stage replication process: find a device, gain code execution, install model software, and continue spreading.
- The experiments used Android emulators with unauthenticated debugging access, which made access easier than on properly secured devices.
- Researchers did not demonstrate a model that could both complete the replication process and run on the targeted consumer hardware.
- Anthropic projects that smaller models could gain the required capability by the end of 2026 if progress continues at its historical rate; the work has not been independently verified.
- Existing botnets such as Mirai and BADBOX 2.0 show that compromised smart devices can already be abused, but neither demonstrates autonomous AI replication.
- Recommended precautions include updating device software and firmware, choosing vendors that provide security updates, avoiding unofficial app stores, and isolating smart devices on a separate network.
Article Details
- Event Type
- Anthropic research warning about the potential for small AI models to self-replicate across vulnerable IoT devices.
- Impact
- The research demonstrated AI agents attempting a replication sequence in Android emulators with unauthenticated debugging access, but did not demonstrate autonomous AI replication on compromised household devices or a model both capable enough and small enough to run on the targeted consumer hardware. Anthropic projects that smaller models could reach the required capability by the end of 2026 if improvement continues at its historical rate. The article notes that existing smart-device botnets, including Mirai and BADBOX 2.0, already abuse connected devices, but do not demonstrate autonomous AI replication.
MITRE ATT&CK
People
Malware
Badbox 2.0used to launch distributed denial-of-service (DDoS) attacks, overwhelming services with traffic. BADBOX 2.0 illustrates another form of abuse: compromised consumer devices used for activities includingMiraiCriminals hijacking connected devices is nothing new. Mirai and its descendants have recruited routers, security cameras, and digital video recorders into botnets used to launch distributed denial-of-service (DDoS)
Vendors
AnthropicResearchers at Anthropic demonstrated this by running a Google Gemma model across 15 off-brand Android TV boxes. They warn that small models could soon spread across vulnerable internet-of-things (IoT) devices,GoogleResearchers at Anthropic demonstrated this by running a Google Gemma model across 15 off-brand Android TV boxes. They warn that small models could soon spread across vulnerable internet-of-things (IoT) devices,
Products
AndroidResearchers at Anthropic demonstrated this by running a Google Gemma model across 15 off-brand Android TV boxes. They warn that small models could soon spread across vulnerable internet-of-things (IoT) devices,GemmaResearchers at Anthropic demonstrated this by running a Google Gemma model across 15 off-brand Android TV boxes. They warn that small models could soon spread across vulnerable internet-of-things (IoT) devices,Google Play ProtectDon’t disable protections to make an app work. Requests to turn off Google Play Protect are a warning sign, as are Android devices without Play Protect certification.