Anthropic Researchers Warn AI Worms Could Self-Replicate Across IoT Networks
Analysts say autonomous AI agents could soon create self-replicating malware that adapts to unfamiliar devices, posing a new threat to internet-connected hardware.
Key points
- Anthropic researchers warn small AI models may soon self-replicate on commercial IoT devices like smart ovens.
- Unlike traditional malware, AI worms could generate tailored attack strategies for each unique target encountered.
- Researchers explored splitting large models into shards to run distributed attacks across low-power machines.
Researchers at Anthropic have identified a growing threat from artificial intelligence agents capable of creating self-replicating worms that can infect any online device. According to Techdirt, these autonomous systems could seek out and hack into unfamiliar hardware without human intervention, marking a significant shift from previous malware capabilities.
The Anthropic team notes that small AI models might be months away from gaining the ability to self-replicate on commercial internet-of-things devices. Targets could include smart ovens and off-brand Android TV boxes. Unlike non-AI malware, such a replicator could figure out unfamiliar devices on its own, potentially bypassing traditional security measures that rely on known signatures.
Historical Context and Evolution
This development echoes the 1988 Morris worm, which spread across early internet networks and shut down thousands of VAX and Sun computers. The Morris worm was an example of software doing exactly what its creator intended, albeit without understanding the consequences of rapid replication.
Since that incident, malicious worms have been designed to create botnets for distributed denial-of-service attacks or blackmail. Examples include Mirai and the later Aisuru-Kimwolf. However, these earlier worms targeted specific kinds of internet-connected hardware and lacked the adaptive reasoning capabilities of modern AI agents.
Technical Mechanisms and Capabilities
Previous research has demonstrated that AI agents can carry out attacks on more powerful systems by generating tailored strategies for each target. The worm parasitically uses compromised machines to run open-weight large language models to sustain its reasoning or extend its reach for further attacks.
Anthropic researchers explored whether an AI worm could run a capable model in a distributed fashion across low-power IoT devices. They investigated splitting a model that is too large for one device into pieces, or shards, that run on separate machines. These shards communicate with each other to collectively run the model.
Who Is Affected
The threat extends to networks spanning Linux, Windows, and Internet of Things devices. Previous work showed that AI worms propagated by exploiting common, real-world corporate network vulnerabilities. The ability to run distributed models on low-power hardware means that even inexpensive consumer electronics could become part of a larger attack infrastructure.
That suggests organizations must prepare for malware that adapts to its environment rather than following a static script. The distributed nature of the attack means that removing a single infected device may not stop the broader campaign if the model shards remain active elsewhere.
What should you do?
Security professionals should monitor for unusual network traffic patterns that might indicate distributed processing across multiple low-power devices. Organizations should review their network segmentation strategies to limit the lateral movement of potential AI-driven threats.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Anthropic
- Monitor network logs for anomalous communication patterns between IoT devices and external servers that might indicate distributed model execution.
- Segment networks to isolate low-power IoT devices from critical corporate infrastructure, limiting the potential spread of self-replicating agents.
- Keep all firmware and software updated to patch known vulnerabilities that AI worms might exploit to gain initial access to devices.
- Review access controls to ensure that compromised machines cannot easily recruit other devices into a distributed attack network.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How soon could AI worms become a reality?
Anthropic researchers suggest that small AI models may be months away from being capable enough to self-replicate on commercial IoT devices.
What is the difference between AI worms and traditional malware?
AI worms can generate tailored attack strategies for each target and figure out unfamiliar devices on their own, unlike static malware.
Can AI worms run on low-power devices?
Yes, researchers explored splitting large models into shards that run on separate low-power machines, allowing collective execution of the model.
