Anthropic has warned that small AI models could soon gain the ability to copy themselves across internet-connected devices without human help. These models could spread like computer worms by exploiting security flaws in connected devices. These devices could include smart TVs, ovens and other household appliances.
Anthropic published a research report on October 12, 2026. The company said small AI models could gain the ability to replicate themselves across Internet of Things (IoT) devices within months. Such systems could be harder to stop if their copies spread across multiple devices instead of relying on a single server.
However, current small AI models still struggle to complete every step reliably. Anthropic’s research points towards a possible cybersecurity risk. However, there is no proof that an AI worm capable of spreading widely is already active in the real world.
Anthropic Tests Whether AI Models Can Replicate Themselves
Anthropic’s Frontier Red Team tested whether AI models could find vulnerable devices, gain access to them, copy their files, and run on other devices without human help.
The researchers focused on internet-connected devices because many consumer Internet of Things (IoT) products have weak security. If an AI model gains access to these devices, it could use their computing power to run elsewhere.
The team tested AI models with 30 billion parameters or fewer to see whether they could perform the tasks needed for self-replication. Several models were able to complete many of these tasks. However, putting all the steps together into a reliable process remained a challenge.
In one simplified test, Anthropic found that the GLM-5.2 model could replicate itself across a test network. It reached seven or eight of the eight target devices in each of five trials.
However, the test used a simplified security setup. The researchers gave the model access to devices through an Android Debug Bridge (ADB) port that did not require authentication. This made it easier to access the test devices than it would be in a properly secured network.
This means the results do not prove that AI models could spread as easily across properly secured devices in the real world.
Small AI Models Could Run on Smart TVs and Other IoT Devices
Anthropic said AI models that can copy themselves may use the computing power of hacked consumer devices instead of expensive data-centre GPUs. Potential targets include:
- Smart TVs and Android TV boxes: These devices could be targeted if they have weak security.
- Smart home appliances: Connected refrigerators, ovens and other appliances could be at risk if attackers find security flaws.
- Security cameras and video-conferencing devices: Some devices have AI chips that can run models directly.
- Small-business IoT devices: Poorly secured devices could provide extra computing power.
A large AI model could also be split into smaller parts and run across several devices that work together. Anthropic tested a 26-billion-parameter model across 15 Android TV boxes. The experiment showed that low-cost devices could work together to run a model that would normally need more powerful hardware.
However, this approach may be slow. Anthropic estimated that the replication process could take days, depending on the model and hardware.
AI Worms Could Spread Across Devices Without a Central Server
Traditional computer worms spread by exploiting security flaws in computers and networks. Anthropic warned that AI-based worms could use AI models to find vulnerable devices and spread with little human help.
An AI worm could find a weakly protected device, copy its files to another machine, and repeat the process. This could help it spread across multiple devices. Taking one infected device offline might not stop the spread, as other devices could still run copies of the worm.
Anthropic also warned that these worms could use different AI models to perform different tasks. However, researchers have not shown that AI worms can reliably bypass modern security systems or spread widely in real-world networks.
ALSO READ: AI Cyber Attacks Statistics [2025 and beyond]
Small AI Models Could Replicate Themselves by the End of 2026
Anthropic predicts that small AI models could gain the ability to replicate themselves by the end of 2026 if current progress continues. The company studied open-weight models with 30 billion parameters or fewer and tracked how their capabilities improved over time.
However, this is only an estimate. Anthropic has not confirmed when self-replicating AI will become possible. The smallest models tested could not yet complete all the steps needed for reliable self-replication. They would need to improve further to reach that level.
How Self-Replicating AI Differs From Traditional Malware
Traditional malware usually targets specific devices or security flaws. Self-replicating AI could adapt to different devices and systems.
Anthropic warned that AI models could learn how to run on unfamiliar hardware, making them harder to stop. They could also learn from failed attempts and share successful methods with other copies.
However, researchers have not shown that AI worms can bypass all security protections or reliably spread across different devices. The tests focused on individual steps, not the full process working in real-world conditions.
ALSO READ: ARTEX AI Agent Pulled From GitHub After Cyberattacks Target South Korean Banks
How Can Companies Prevent AI Worms From Spreading?
Anthropic said companies can reduce the risk by securing connected devices and monitoring their networks.
- Secure devices: Change default passwords, close unused access points and install security updates.
- Monitor networks: Watch for unusual connections, large file transfers and suspicious activity.
- Limit access: Restrict what devices can access to reduce the damage from attacks.
- Detect malware early: Look for signs of automated scanning and malware spreading across devices.
- Build stronger security: Manufacturers should include better security protections in their products.
Anthropic also recommended limiting how many new devices an infected machine can reach. If each infected device spreads the threat to fewer than one other device on average, the spread should eventually stop.
