Anthropic has published a controlled study warning that increasingly capable AI models could eventually help power self-replicating malware, sometimes described as an “AI worm,” across poorly secured internet-connected devices. The report is a research warning about a possible capability—not evidence that such an AI worm is currently spreading in the wild.
What did Anthropic test?
The researchers studied whether open-weight models could complete parts of a replication chain: identify a reachable machine, establish an environment to run a model, and continue to another host. Their experiments included a firewalled fleet of Android 15 emulators. In that simplified setup, devices were deliberately reachable over an unauthenticated debugging port, so the test did not reproduce the full difficulty of breaking into ordinary protected devices.
Anthropic also demonstrated distributed inference using Gemma 4 26B across 15 inexpensive Android TV boxes, at about 0.39 tokens per second. That shows that a model can run slowly across low-cost hardware; it does not show that the boxes were infected or that an autonomous attack was launched on the public internet.
What did the results show—and not show?
In Anthropic’s simplified end-to-end evaluation, GLM-5.2 was the only tested open-weight model reported to complete the full replication process reliably. The company says the exploit step was made trivial, which means the result is an upper-bound demonstration of the model’s ability under those test conditions, not a measure of real-world attack success.
Anthropic estimates that if small-model capabilities keep improving at the historical rate it measured, a model could reach that capability level by the end of 2026. The authors describe this as a rough projection based on a proxy metric, not a confirmed deadline or prediction that an outbreak will happen.
Why smart devices matter
Internet-connected televisions, cameras, routers and appliances can be difficult to secure when they use weak default credentials or stop receiving updates. Anthropic argues that a future automated threat could attempt to adapt across different hardware. The study did not demonstrate a real-world infection campaign, and it did not show that common smart-home products are automatically vulnerable.
Practical steps for households
CISA’s digital-home guidance recommends changing default router and device passwords, keeping software and firmware updated, enabling automatic updates where available, and reviewing app permissions. These steps reduce familiar risks from existing cyber threats as well as potential future ones.
Image note: conceptual cybersecurity illustration; it does not depict Anthropic’s experiment. Image: jaydeep_ via Wikimedia Commons, released under CC0.
