TVs, appliances could become launchpads for AI-driven attacks: Anthropic

Anthropic research finds small AI models can perform steps needed for self-replication across devices. Experts are of the view that these early warnings should not be taken lightly

Anthropic, in an in-house experiment, found that small artificial intelligence (AI) models can perform many of the steps needed to replicate themselves across connected devices. Such capabilities could allow compromised smart TVs and appliances to become part of a network used to spread malware or run AI models. Such a network could also be difficult to shut down because it would not depend on a single central server.

 

Small AI models could power future cyberattacks

 

Anthropic found that AI models with 30 billion parameters or fewer could carry out many of the individual steps needed to copy themselves across devices, although they struggled to complete the entire process reliably.

 

 

In one experiment, the researchers tested GLM-5.2 on a simulated network of eight devices. The model reached all eight devices in four out of five trials and seven devices in the fifth. The researchers also tested whether smaller models could run on different hardware. A model with nine billion parameters managed to run AI inference on eight of the nine hardware types tested.

 

Anthropic said that if models continue to improve at their previous rate, smaller AI models could become capable of reliably replicating themselves by the end of 2026. However, this is the researchers' projection, not a confirmed timeline for such attacks to emerge in the real world.

 

Experts are of the view that these early warnings should not be taken lightly.

 

“We have not yet seen a fully autonomous AI worm spreading across real networks. But the gap could close faster than expected as smaller models become more capable. The research is best read as an early warning and as a reason to fix the basics now,” said Jaydeep Singh, general manager for India at Kaspersky. 

 

Why smart TVs and appliances matter

 

Anthropic's research raises concerns that attackers may not need expensive data centres or powerful chips to run AI models. In one experiment, researchers ran a model called Gemma4-26B-A4B across 15 inexpensive Android TV boxes, each with 4 GB of memory. They split the model into smaller parts so the devices could work together to run it.

 

If attackers gain control of enough connected devices, they could potentially combine their computing power to run AI models. Smart TVs, cameras and appliances could become targets if they have weak security. Such devices are often left switched on and may not receive regular software updates, making them attractive to attackers.

 

Anthropic pointed to Kimwolf, an existing internet-of-things (IoT) botnet that infected more than two million devices, mainly off-brand Android TV boxes.

 

“Cheap connected devices are always on, rarely monitored and seldom updated, which is why botnets have relied on them for years. Mirai variants still make up most of the threats we see targeting IoT devices,” said singh. He added that attackers would likely exploit familiar weaknesses, including default passwords, exposed services such as Telnet or SSH, and outdated software.

 

How could AI change the way malware spreads

 

Conventional malware can copy itself across vulnerable systems, but it generally follows instructions written by its creators. Anthropic's research examines whether AI models could help automate the process of identifying unfamiliar devices and adapting to different hardware.

 

The researchers also tested a method in which one AI agent passed notes about its attempts to a fresh agent. Later agents could use this information to try different approaches instead of repeating earlier mistakes.

 

In these experiments, the success rate for setting up AI inference on unfamiliar hardware increased from 45 per cent in the first generation to between 61 per cent and 84 per cent in later generations. This does not mean the model was independently retraining itself. Rather, later agents benefited from information shared by earlier ones.

 

Singh said conventional malware follows a script, so it usually fails when it lands on a device it was not written for. An AI-driven worm could study its environment, identify the system it is dealing with and adjust its approach, whether that involves exploiting a vulnerability, installing itself or staying hidden.

 

This flexibility could make it harder to detect, as signature-based tools rely on recognising known code and behaviour. However, AI models are much larger than typical malware, which can slow their movement across networks and make them easier to spot.

 

What should enterprises do now

 

For Indian enterprises, connected devices, cloud services and AI can no longer be secured as separate projects. A CCTV camera, a smart TV in a meeting room or an AI agent with broad permissions can each become a way in.

 

According to Singh, current practices cover part of this risk, but most are built around known malware and human-led attacks. Many organisations still do not have a full inventory of their IoT devices and AI agents, and that is where the work has to start.

 

Singh said that device manufacturers should get rid of default passwords, ship devices with unused services switched off and support firmware updates for the device's lifetime. They also need to secure their own supply chain, as backdoors can be built into firmware before a device is sold.

 

Enterprises, meanwhile, should change default credentials, update firmware, isolate devices on separate networks and turn off remote access where it is not needed. For AI agents, the guidelines recommend treating external data as untrusted and requiring human approval before any irreversible action.

 

“Traditional cybersecurity models are built around a human threat actor. But when code can learn, adapt, conceal itself, and potentially replicate, security must evolve from signature-based detection to behaviour-based containment,” said Ashish Sharma, Partner at Deloitte India.

 

Sharma said organisations need to strengthen multiple layers of defence, including identity and privilege management, network ingress and egress controls, EDR/XDR systems calibrated to detect AI-driven behaviours, robust governance frameworks and, where appropriate, a reliable 'kill switch'.