AI model ‘self-replication’ is only months away, Anthropic says — and it doesn’t even need a data center to happen

Anthropic, an AI research company, has warned that self-replicating artificial intelligence (AI) could become a reality within the next few months. These so-called AI "worms" could spread across various hardware devices, including Internet of Things (IoT) devices and data centers, without the need for a centralized off-switch to control them.

Smaller AI models are already able to adapt to unfamiliar hardware, and self-replication is possible. For instance, Anthropic's researchers tested this concept using a 25 billion parameter model (Gemma4-26B-A4B) across 15 inexpensive off-brand Android TV boxes, which cost around $55 each. Once networked together, the TV boxes were able to run distributed inference, proving the possibility of an AI worm.

IoT devices are a prime target for these worms because they are commonly "cheap and poorly defended." Self-replicating AI models could exploit weak cyber defenses of IoT devices in homes and small businesses, split their model into "shards" across the compromised hardware, and then pay for their own token costs by offering their inference on a digital freelancing service.

Larger and more powerful AI models would also pose a significant threat. These worms could breach hardened defenses of datacenters and other high-compute facilities, using the hardware to propagate further. An additional threat is an AI worm that uses an API-based replicator, which steals human credentials to access commercial AI providers and use their services to pay for their own inference.

Anthropic notes that self-replicating AI isn't an immediate threat, but it could become a problem by year-end. One major hurdle for these worms is operating on unfamiliar hardware. Unlike traditional malware, AI worms could adapt to new environments and hardware. Smaller models, such as Qwen3.5-9B, are already capable of adapting to unfamiliar hardware, successfully running inference on 8 out of 9 tested hardware types, including Jetson Nanos, Mac PCs, and RTX 4090s.

Defenders have a potential advantage over AI worms. To balance stealth and replication speed, AI worms must leave a noticeable network signature, such as multi-gigabyte model payloads delivered to newly infected devices or unusual traffic between machines. Network quarantine could be an effective initial defense, and implementing a whitelist of approved sites could help clear AI worm infections.

Regulations on IoT device security, such as the UK's decision to move away from supplying devices with default passwords, could also help prevent AI worms from gaining a foothold.