Anthropic has warned that small AI models may soon be able to copy themselves onto poorly secured smart devices without needing a data centre. The company’s Frontier Red Team tested whether a model could find vulnerable devices, install itself, run there, and keep spreading without human help. Researchers bought 15 off-brand Android TV boxes for about $55 each and ran Google’s open model Gemma4-26B across them, though performance was slow at less than half a token per second. Open models with 30 billion parameters or fewer can already handle many individual steps but still make mistakes when completing the full sequence. In a simplified test where the hacking step was made easy, only the larger open model GLM-5.2 managed end-to-end spread across a small network of emulated Android devices. Anthropic said that if small models keep improving at their current pace, they could reach this capability by the end of the year.
Anthropic has warned that small AI models may soon be able to copy themselves onto poorly secured smart devices without needing a data centre. The company’s Frontier Red Team tested whether a model could find vulnerable devices, install itself, run there, and keep spreading without human help. Researchers bought 15 off-brand Android TV boxes for about $55 each and ran Google’s open model Gemma4-26B across them, though performance was slow at less than half a token per second. Open models with 30 billion parameters or fewer can already handle many individual steps but still make mistakes when completing the full sequence. In a simplified test where the hacking step was made easy, only the larger open model GLM-5.2 managed end-to-end spread across a small network of emulated Android devices. Anthropic said that if small models keep improving at their current pace, they could reach this capability by the end of the year.