Anthropic's AI worm warning faces skepticism over hardware limits

The AI worm hypothesis

Anthropic recently published a report titled “AI models may be capable of self-replicating without access to a datacenter.” The company warns that small AI models are advancing quickly enough to colonize internet-connected consumer electronics. This could potentially result in decentralized botnets that are difficult to shut down because they lack a single point of failure.

The report suggests that everyday items—including smart bulbs, ovens, fridges, and even cars—could become hosts for these self-replicating programs. Anthropic uses the term "AI worms" to describe software that can autonomously find new machines and copy itself across a network. This scenario paints a picture of a "digital slime mold" spreading through the modern smart home.

The company's research aims to show that these threats are not just theoretical but could emerge within months. They argue that as models become more efficient, the need for massive data centers to run them will decrease. This shift could allow malicious actors to deploy intelligent, self-propagating code on a massive scale using existing consumer hardware.

Testing the feasibility

To test the possibility of distributed AI, Anthropic researchers attempted to run Google’s Gemma 4 model across 15 off-brand Android TV boxes. Each of these boxes cost about $55. The combined setup produced an inference speed of 0.39 tokens per second, which equates to roughly 18 words per minute.

The researchers also simulated a network of Android devices to test an AI-powered worm using the GLM-5.2 model. This specific model has 744 billion parameters and is over 1TB in size. In four out of five trials, the system managed to replicate across eight different nodes in the simulated environment.

However, the mechanics of this test revealed a significant caveat. The large GLM-5.2 model did not actually replicate itself onto the target devices. Instead, it functioned as an external inference server that set up a much smaller model, Llama 3.1 8B, on each of the target machines. This means the replication was orchestrated by a large, externally hosted model rather than a small model running independently on the hardware.

Physical hardware constraints

Critics of the Anthropic report point to the "size problem" as a major hurdle for AI worms. AI models are not magic; they are massive files that require significant memory to run. For example, a 26-billion-parameter model like Gemma 4 requires between 16 and 20 gigabytes of disk space even when compressed.

Most IoT devices are severely underpowered compared to the requirements of modern AI. A smart fridge might only have a few hundred megabytes of RAM, while a smart TV box might have 4GB. Even a "small" 8-billion-parameter model is still several gigabytes in size, making it nearly impossible to fit into the tiny chips found in devices like Roombas or smart lightbulbs.

To run a model on these devices, a worm would have to distribute the model across dozens or even hundreds of coordinated machines. Even if this were achieved, the inference speed would be so slow that the worm's intelligence would be functionally useless. As Sify News reported, the physical reality of these hardware limits makes the scenario conceptually possible but physically next to impossible.

Real threats vs. hype

Many cybersecurity experts argue that the fear of AI worms may be a form of vendor exaggeration. Ciaran Martin, the former head of Britain’s National Cyber Security Centre, noted that running a small model on cheap TVs does not necessarily constitute a security crisis. He suggested that the report might be designed to generate headlines rather than reflect an immediate danger.

The reality is that traditional malware, like the Morris worm from 1988, is still a very effective threat. Those older viruses were tiny, taking up only kilobytes of space, which allowed them to be fast and deadly. They do not need to "think" or carry a large AI brain to exploit vulnerabilities in poorly secured systems.

The actual risk may be that AI is making existing cyberattacks cheaper and more convincing. Agencies like CERT-In have already warned about AI-generated phishing and the rapid weaponization of vulnerabilities. The devices Anthropic lists as vulnerable are at risk not because they can run AI, but because they are cheap, rarely updated, and often use unencrypted protocols.

The path forward

While the AI worm apocalypse may be a distraction, the underlying trends are worth watching. The surface area of attack is expanding as every connected device becomes a potential entry point for hackers. The convergence of AI tools and the proliferation of IoT devices creates a landscape where basic security hygiene is more important than ever.

The Anthropic report serves as a reminder that the industry is moving toward a more complex threat environment. However, experts suggest that focusing on the "doomsday" scenario can lead to panic while ignoring the problems already in existence. The most immediate dangers are deepfake voice scams, convincing phishing emails, and the manipulation of AI agents with access to internal systems.

Ultimately, the machines are not likely to take over home appliances in the way the report suggests. The real fight is about securing the massive expansion of connected devices. We must ensure that as AI makes attacks faster and more convincing, our fundamental security practices keep pace with the technology.

What to watch

Watch for how AI models continue to shrink in size while maintaining high performance. If models can eventually fit into the RAM of common household appliances, the feasibility of the AI worm scenario will increase significantly.

Monitor official advisories from agencies like CERT-In regarding the weaponization of AI in phishing and multi-stage attack orchestration. These represent the most immediate and documented risks to both individuals and corporations.

Try it

Check your home router's security settings to ensure that all connected IoT devices are using encrypted protocols and have the latest firmware updates installed.