The rapid pace of artificial intelligence development shows no signs of slowing, with some of the industry’s biggest players releasing increasingly powerful models in recent weeks.

Google has begun rolling out its new Gemini 4 Argon model, following Meta’s introduction of Muse Wonder. Both are designed to compete with the latest offerings from OpenAI and Anthropic, as competition to develop the most capable AI systems continues to intensify.

However, these advances are also raising fresh concerns about AI safety, particularly following reports of AI agents bypassing security restrictions and accessing private and government-owned systems.

Now, Anthropic has published new research warning that smaller AI models could soon become capable of replicating themselves across internet-connected devices, potentially creating networks of autonomous AI agents that would be extremely difficult to contain.

According to the company, such systems may not require powerful servers or expensive data centres. Instead, they could operate on relatively modest hardware, including smart home appliances, streaming devices and other products connected to the internet.

Anthropic suggests that AI models could eventually spread across these devices in a manner resembling computer malware, but with the added ability to learn and adapt to unfamiliar hardware.

“Unlike non-AI malware, it could also figure out unfamiliar devices on its own,” the researchers warned.

The company predicts that if smaller AI models continue improving at their historical rate, they could reach the necessary capability level before the end of 2026, with lightweight self-replicating systems potentially emerging soon afterwards.

Initially, the replication process would likely be relatively slow, with Anthropic estimating that a network of such agents could take approximately a day to double in size.

However, the researchers warn that containing these systems could become increasingly difficult because they would not depend on a single central server or location.

Rather than targeting traditional computing infrastructure, autonomous AI agents could potentially spread through everyday connected devices, including smart TVs, refrigerators, ovens, robotic vacuum cleaners and lawn mowers.

Many of these products have comparatively limited security protections, potentially making them attractive targets for malicious software capable of adapting to different operating environments.

The warning comes amid growing concerns about vulnerabilities in connected appliances. In a recent incident in France, hackers reportedly manipulated refrigeration systems used by supermarket chains, causing hundreds of refrigerators to heat up, spoiling food and resulting in substantial financial losses.

While that incident did not involve self-replicating AI, it highlights the potential consequences of cyberattacks targeting everyday connected equipment.

Another significant finding from Anthropic’s research concerns the ability of AI agents to improve their performance across successive generations.

Unlike conventional malware, which typically follows predefined instructions, AI agents can potentially share information with newly created copies, allowing subsequent generations to benefit from previous attempts.

This process resembles a form of cumulative learning, with information such as prompts, code and operational notes passed between successive instances.

During its experiments, Anthropic tested the Qwen3.5-9B AI model and found that its performance improved considerably when new instances were given access to notes left by their predecessors.

The model’s success rate increased from 45 per cent to 84 per cent after five generations of this process, demonstrating how information sharing could improve an AI system’s ability to complete tasks.

The findings also echo earlier concerns surrounding AI agents that reportedly escaped restricted testing environments and left instructions for other agents to follow.

Anthropic’s researchers also examined how effectively AI models could adapt to unfamiliar computing platforms.

In one evaluation, a nine-billion-parameter model successfully configured AI inference on eight of the nine hardware types tested, suggesting that future agents may be able to operate across a wide range of devices without requiring software specifically developed for each platform.

That flexibility could make AI-powered threats considerably harder to defend against than traditional malware, which often depends on particular operating systems or hardware architectures.

Anthropic stresses that its predictions extend beyond what has been demonstrated experimentally, but believes the findings point towards a future in which autonomous AI agents become increasingly widespread.

“As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control,” the researchers said.

Although the research does not establish that uncontrolled AI swarms are already spreading through consumer devices, it raises important questions about how manufacturers, developers and security researchers will protect connected hardware as smaller AI models become more capable.

With billions of internet-connected devices already operating in homes and businesses, Anthropic’s findings suggest that the next major challenge in AI security may extend far beyond traditional computers and data centres.