Anthropic conjures another threat that doesn't exist yet to stoke fears of AI apocalypse

Anthropic, a prominent AI research company, has released a new study claiming that self-replicating AI worms could pose a significant threat to Internet of Things (IoT) devices within the next few months. In a research report published on Monday, the lab suggested that AI models may eventually be able to replicate themselves without relying on datacenter hardware.

The study demonstrated a 9-billion-parameter model successfully setting up inference on eight out of nine hardware types tested by Anthropic. The researchers noted that such a self-replicating AI botnet would be difficult to shut down due to the lack of a centralized point of control. Moreover, the AI worm could potentially adapt more easily to unfamiliar devices compared to traditional botnets like Mirai and Kimwolf.

To test their hypothesis, Anthropic's researchers managed to run a distributed instance of the Gemma4-26B-A4B model across 15 generic Android set-top streaming boxes, which are often targeted by traditional botnets due to their poor security. Once operational, the distributed Gemma4 model was able to perform inference tasks at a rate of 0.39 tokens per second.

The researchers then utilized Z.ai’s GLM 5.2 to propagate a copy of Llama3.1-8B across eight Google Cloud instances running an Android 15 emulator in a virtual private cloud. The model successfully spread to all eight nodes in four out of five trials and seven of eight nodes in the remaining trial. Anthropic acknowledged that the replication process demonstrated in their study does not include the cyber exploit element, as earlier papers have already shown that small models can carry out cyber exploits with the right tools.

However, they also noted that IoT devices are often insecure, but good security hygiene could mitigate most common ingress routes. The speed at which an AI botnet could grow was also considered, with Anthropic estimating that it might take days for the botnet to double in size, compared to Mirai's ability to double in around an hour.

While the study showed that AI models can complete many individual steps required for replication, end-to-end replication has not been achieved yet. Anthropic further emphasized that the future of autonomous AI replication is uncertain and that basic security measures could help prevent such threats from arising.