Key highlights:
- Anthropic’s researchers are studying if small AI models might start copying themselves across internet connected devices at home, even if people don’t get involved much.
- They ran an experiment with 15 cheap Android TV boxes and showed that AI can run on affordable consumer gadgets.
- In a controlled test, GLM-5.2 reached all 8 Android emulator nodes in most trials, but the setup did not have strict security barriers.
- Researchers think smaller AI models could get smart enough to self replicate by the end of 2026, judging by how fast AI has improved so far.
Anthropic’s Frontier Red Team released new research looking into whether AI models could start replicating across devices like smart TVs and other home electronics. The study asked if smaller AI models can spot new types of hardware, set up AI inference, and figure out steps to jump from one device to another.
The team wrote about progress in teaching AI to act on its own, but also pointed out technical challenges they haven’t solved yet. The experiments happened in closed lab settings, so they did not actually unleash an AI worm into real networks. Mostly, their findings show where AI is heading and what this might mean for keeping devices secure.
Researchers Test AI Models on Low-Cost Smart TVs
One experiment used Google’s Gemma4-26B-A4B model running across 15 lesser known Android TV boxes. Each box cost around $55 and had 4GB of memory. By splitting different parts of the model across the 15 boxes, they got an inference speed of 0.39 tokens per second. This test checked if cheap home electronics could team up to provide enough power for an AI model. Importantly, they did not test spreading a worm across these devices.
They also tried to see if AI models could set themselves up on hardware they’d never seen before. That meant the AI had to pick the right software, install what it needed, and fix tech issues that came up from different operating systems or hardware. A model with 9 billion parameters got inference running on 8 out of 9 device types, including Jetson Nano boards, Macs, and RTX 4090 graphics cards. Basically, the models showed they could adapt to a lot of setups, though getting inference to work is just one step in actually copying themselves.
AI “worms” could soon be capable of spreading across internet-connected household devices, with little human intervention, according to new research from Anthropic. “With much less human intervention, you could be looking at potentially destructive acts happening in environments that weren’t predicted,” cybersecurity expert Jordan Rae Kelly says.
— CBS News (@CBSNews) October 13, 2026
Controlled Network Tests Show Mixed Results
To see if these models could handle more complex tasks involved in replication, Anthropic set up a private network of 8 Google Cloud nodes, each running an Android 15 emulator. The experiment tested whether an AI agent could spot new machines, get access, send over its payload, set up AI inference, and keep the cycle going. GLM-5.2 reached all 8 nodes in 4 out of 5 tests, and in the last one, made it to 7 nodes.
But they used an Android Debug Bridge without authentication, so access was easier, most real devices would be better protected. There’s another catch: the controlling model ran on its own infrastructure, while the target machines ran Llama3.1-8B. So they did not show that GLM-5.2 could copy and run itself on everyday home electronics by itself.
Anthropic said that AI models with up to 30 billion parameters can do lots of individual replication tasks, but often struggle when trying to chain those steps together reliably. Looking back at how much open-weight models have improved, the researchers guessed that these smaller models could be ready to do full replication by the end of 2026. But that’s just an estimate, not a confirmed deadline for the arrival of a working autonomous worm.
Experiments Examine Replication Speed and Adaptation
The study also looked at how long it takes to set up inference on devices that aren’t very powerful. In their tests, Qwen3.5-9B needed between 28 and 51 hours to finish its setup. Gemma4-26B-A4B, which wasn’t optimized, took even longer: 71 to 150 hours.
That means if an AI worm used similar devices, it would copy itself much slower than regular malware, taking days, not hours, to double. They also wanted to see if these AI agents could learn from earlier tries. With a system called Memento, new Qwen3.5-9B runs got notes about what previous runs did right or wrong.
Their inference-setup success rate increased from 45% in the first generation to between 61% and 84% in later generations. The improvement came from passing information between instances, not from retraining the model’s main brain.
Researchers also tested if getting help from a bigger, smarter model would make a difference, but saw no clear improvement for these tasks, at least, the smaller models did not benefit much from unfamiliar devices.
Researchers Outline Possible Cybersecurity Measures
Anthropic suggests a few ways to help stop future AI worms: monitoring networks, improving device security, and making sure teams work together during outbreaks. Moving big AI model files around could make devices behave strangely like exchanging large amounts of information, or talking to other devices they normally ignore. Those could be warning signs for security teams.
The report also talks about what worked against the Mirai botnet. In one example, quarantining and warning users helped wipe out 92% of infections in 14 days, versus just 74% if users weren’t notified. Other fixes include stronger logins, safer default settings, and better built-in security from device makers. To be clear, the study isn’t saying that AI worms are attacking homes or companies right now. Instead, it shows how well models can handle certain steps, what they still struggle with, and what advances in smaller AI models might mean for future device security.
