AI model ‘self-replication’ is just months away, says Anthropic, and you don’t even need a data center to make it happen
- Anthropic warns that self-replicating AI worms could be feasible by the end of the year
- These AI worms could hijack hardware ranging from IoT devices to data centers
- The smallest models are already capable of adapting to unknown hardware and self-replicating.
Anthropic has warned that self-replication could be feasible for small AI models in Internet of Things (IoT) devices in just a matter of months.
This could be a big problem because a self-replicating AI model does not have a centralized kill switch and could jump networks to discover other IoT hardware to replicate it further without human control.
Anthropic warns that these AI worms could use multiple compromised IoT devices to run inferences while carrying the weights of their model with them as it plays. IoT devices would be a prime target for this worm because they are “cheap and typically poorly defended,” the company says.
Latest videos ofTechnologyRadar
Small AI models could spread quickly
To test the feasibility of a self-replicating AI worm, Anthropic researchers ran Gemma4-26B-A4B, a 25 billion parameter model, on 15 inexpensive, unbranded Android TV boxes at a cost of about $55 each. Once networked, the TV boxes were able to run distributed inference, solidifying the proof of concept.
This introduces a number of novel opportunities for small AI models. Self-replicating AI models could exploit the weak cyber defenses of IoT devices in homes and small businesses, and then split their model into “chunks” across compromised hardware to run the model collectively.
The AI worm could then pay its own symbolic cost by offering its inference in an independent digital service.
Additionally, newer IoT hardware that includes embedded AI chips designed to run small models (such as security cameras and video conferencing kits) would be a prime target for an AI worm. Hackers could breach these systems through social engineering or even using an artificial intelligence agent in the not-too-distant future.
Anthropic further warns that data centers and other high computing facilities could become targets for advanced replicators that are capable of breaching strengthened defenses and then using the hardware to spread further.
An additional threat that Anthropic warns about is an AI worm that uses an API-based replicator, whereby the model does not host itself, but rather steals human credentials to access commercial AI providers and use these services to pay for its own inference.
AI self-replication is not a reality…yet
While self-replicating AI is not an immediate threat, researchers theorize that it could become an issue by the end of the year.
One of the main obstacles for these AI worms is becoming operational on unknown hardware. Unlike traditional “worm” malware that must be created to attack specific IoT vulnerabilities, AI worms could be able to adapt to unknown environments and hardware.
In Anthropic testing, researchers found that smaller models, like the Qwen3.5-9B, are already capable of adapting to unknown hardware to run inference. Qwen3.5-9B matched 8 of the 9 types of hardware the researchers tested, including Jetson Nanos, Mac PCs, and RTX 4090s.
While smaller models are better able to adapt to unknown hardware, larger models are better equipped to self-replicate quickly.
Defend against AI worms in the wild
There is a significant trade-off for AI worms, Anthropic says, which works in the defender’s favor. AI worms must balance stealth and replication speed. To that end, researchers say an AI worm would leave a noticeable network signature, such as multi-gigabyte model payloads delivered to newly infected devices or unusual traffic between machines that wouldn’t normally communicate.
Network quarantine would also be an effective initial defense against AI worms. Using lessons learned from the Mirai botnet remediation, researchers believe that implementing a whitelist of approved sites with browsers redirected to cleanup instructions could help eliminate AI worm infections.
Regulations on the security of IoT devices could also help bolster the defense against AI worms. Stopping providing default passwords to IoT devices, as has already been done in the UK, would help prevent AI worms from taking hold. Other standards, such as automatic updates and extended update duration, can help protect IoT devices against traditional malware and AI worms.
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds.
Post Comment