AI Self-Replicating Malware Could Make Botnets More Dangerous, Anthropic Warns

AI self-replicating malware could become a serious cybersecurity concern as artificial intelligence continues to advance. New research from Anthropic’s Frontier Red Team warns that relatively small AI models may eventually be capable of moving between vulnerable devices, establishing themselves on new systems, and helping create autonomous botnets.

The possibility raises concerns about a new generation of cyberattacks. Unlike traditional malware, which often relies on instructions written for specific operating systems and hardware, AI-powered malware could potentially adapt its behaviour when it encounters unfamiliar environments.

However, cybersecurity researchers say the threat should be understood in context. Many of the techniques involved already exist in conventional malware. The major change could be AI’s ability to automate tasks that previously required considerable human expertise.

Anthropic Warns of AI-Powered Botnets

Anthropic’s Frontier Red Team has examined how smaller AI models could potentially help malware spread through vulnerable devices. The research describes a scenario in which AI systems identify compatible environments, establish the infrastructure needed to run models, and communicate with other compromised devices.

If these capabilities become practical outside controlled experiments, attackers could potentially use inexpensive devices such as routers, internet-connected cameras, and other IoT equipment as part of distributed networks.

These devices are attractive targets because many remain connected to the internet for long periods and are not always maintained with regular security updates.

Anthropic’s warning suggests that AI could make it easier to adapt malicious software to different devices without requiring attackers to manually develop every variation.

Still, the research does not establish that fully autonomous AI worms are already spreading across real-world networks.

What Makes AI Malware Different From Traditional Botnets?

Botnets are not a new cybersecurity threat. They are networks of compromised devices controlled by an attacker, often without the owners knowing their equipment has been infected.

One of the best-known examples is the Mirai botnet, which emerged in 2016. Mirai infected internet-connected devices, including routers and security cameras, by exploiting weak or default passwords.

The network was later used in a major distributed denial-of-service (DDoS) attack that disrupted services operated by Dyn, affecting access to popular websites and online platforms.

Traditional botnets generally depend on known vulnerabilities, predefined instructions, and malware designed to work with particular device configurations.

AI could change part of that process. Instead of relying entirely on predefined instructions, an AI-assisted system might be able to analyse an unfamiliar environment and help determine what actions are possible.

That could reduce the amount of manual work required to adapt malware to different hardware and software configurations.

Researchers Question Whether AI Worms Are Truly Self-Replicating

Despite the concerns, experts interviewed by Engadget caution against describing Anthropic’s demonstration as a fully self-replicating computer worm.

Michael Bailey, a professor at Georgia Tech’s School of Cybersecurity and Privacy, argues that the research demonstrated some interesting capabilities but did not convincingly establish the kind of autonomous propagation associated with traditional computer worms.

In particular, he distinguishes between copying an entire malicious program onto another device and installing a smaller component that allows the device to communicate with a central system.

The latter approach could resemble a malware dropper, which compromises a system and installs additional malicious software.

Bailey also points out that adaptability is not entirely new. Existing malware has already operated across different device types and hardware architectures. AI’s potential contribution is automating some of the work required to support those environments.

Chris Canal, CEO of EquiStamp, also believes the technology could become a real concern in the near future. However, the exact capabilities and consequences remain uncertain.

The distinction matters: AI could make botnets more flexible without necessarily creating a completely new category of malware.

Why AI Could Change the Economics of Cybercrime

One of the most important implications may be the cost of developing and maintaining cyberattacks.

Building malware that works across different systems can require technical expertise, testing, and continuous adaptation. If AI models can automate parts of that process, smaller or less experienced criminal groups could potentially gain access to capabilities that previously required more resources.

This could shift the balance between attackers and defenders.

Instead of introducing entirely new attack techniques, AI may make existing techniques easier to reproduce, modify, and deploy at greater scale.

Anthropic’s research also raises questions about open-weight AI models. Once model weights are publicly available, developers and security researchers have less control over how those models are modified or used. Removing safeguards from a model can make it difficult to rely on built-in restrictions as the primary defence against malicious use.

How Can Organizations Defend Against AI-Powered Malware?

Despite the advanced technology involved, many of the recommended protections are familiar cybersecurity practices.

Organizations and device owners should focus on reducing the number of systems that attackers can compromise in the first place.

Important measures include:

• Update devices regularly: Install firmware and security updates for routers, cameras, servers, and other connected equipment.
• Replace default passwords: Use strong, unique passwords and enable multi-factor authentication wherever supported.
• Isolate vulnerable devices: Keep IoT equipment on separate network segments instead of allowing unrestricted access to business systems.
• Monitor unusual activity: Watch for unexpected network connections, abnormal traffic, and devices communicating with unfamiliar destinations.
• Remove unnecessary internet access: Disable exposed services and remote administration features that are not required.
• Maintain an incident response plan: Prepare procedures for isolating infected devices and restoring systems safely.

Internet service providers may also play a role by identifying infected devices and restricting their access to potentially harmful destinations while customers resolve security problems.

Regulation can help as well. The United Kingdom’s Product Security and Telecommunications Infrastructure (PSTI) framework includes security requirements for consumer connectable products, including restrictions on easily guessable default passwords.

To reduce the risk of compromised connected devices, CISA recommends applying security updates, changing default passwords, and limiting unnecessary network exposure.

AI May Accelerate Old Cybersecurity Threats

Anthropic’s warning highlights a growing challenge for cybersecurity: artificial intelligence does not need to invent completely new attack methods to make existing threats more dangerous.

If AI systems can automate vulnerability discovery, adapt malicious code, and help coordinate compromised devices, the development and operation of botnets could become more accessible.

But the research should not be interpreted as proof that unstoppable, self-replicating AI viruses are already a reality. Experts still question how effectively the demonstrated techniques would work against properly secured devices and real-world networks.

For businesses and individuals, the practical lesson remains straightforward. Keep devices updated, use strong authentication, isolate vulnerable equipment, and monitor network activity.

As AI capabilities advance, basic security practices may become even more important—not less.