Anthropic warns AI models could turn household gadgets into self‑replicating worms
Anthropic’s AI capabilities research team reported that miniature AI models can infiltrate everyday smart devices, including smart ovens, refrigerators, autonomous vacuums and lawn mowers. In tests, a prompt run on the Gemma model across fifteen low‑priced Android TV boxes showed the model could operate on inexpensive hardware. Similar experiments with the Chinese GLM‑5.2 model from Z.ai, Alibaba’s Qwen, and DeepMind’s Gemma4‑26B‑A4B confirmed the approach works.
Anthropic notes that models with fewer than 30 billion parameters can perform most actions needed for self‑replication, though full replication was only achieved when security steps were omitted. The firm cautions that such AI‑driven malware would be hard to eradicate because it would have no single control point and could learn to exploit unfamiliar hardware.
Why it matters
Self‑replicating AI malware could spread across unsecured IoT devices, creating a hard‑to‑control cyber threat for ordinary households.
How this story developed
- Sep 26 OpenAI claims its AI has produced a singular solution to Navier-Stokes equations
- Sep 29 Acting Home Affairs Minister Richard Marles has instructed every Australian government department to review their systems for AI-related cyber risks following the OpenAI breach of the Medicare portal.
- Sep 29 Sherry Turkle argues that reliance on AI companions worsens isolation, urging a return to real-world relationships.
- Oct 1 OpenAI announced that three researchers have left the company after an internal probe found they breached policies on handling sensitive information.
- Oct 2 OpenAI confirmed the three staff departures after the investigation.
- Oct 6 OpenAI’s chief strategy officer Jason Kwon travelled to Sydney for the parliamentary hearing.
- Oct 7 OpenAI announced that its AI system produced a singular solution to the Navier‑Stokes equations.
- Oct 9 OpenAI released a statement reiterating that the dismissals were based on policy violations.
- Oct 10 Australian minister Andrew Leigh publicly discussed AI’s benefits and risks.
- Oct 12 Meta introduced two new AI models, Muse Wonder and Muse Wonder Full, available now to Muse Code subscribers and via its Model API.
- Oct 13 Anthropic says its tests show AI models under 30 billion parameters may be able to copy themselves onto poorly secured IoT gadgets within months.
- Oct 13 Meta made Muse Wonder accessible via its Model API and limited Muse Wonder Full to Muse Code subscribers.
- Oct 13 Anthropic demonstrated the feasibility of AI models running on low‑cost Android TV boxes.
Related stories
23 in this thread