Anthropic says small AI models could soon copy themselves onto everyday devices
Small AI models may be just months away from being able to copy themselves across everyday household devices, such as smart ovens and TV boxes, according to new research from Anthropic.
TL;DR
- Anthropic's research indicates small AI models may soon be capable of self-replication across smart devices.
- This potential for self-replication raises concerns about a new era of 'rogue AI' spreading like computer worms.
- Current small AI models can perform many steps of self-replication but cannot yet complete the entire process independently.
- Anthropic predicts the gap in capabilities will close by the end of the year, with models matching current successful self-replicators by the end of 2026.
- Small, open-weight models are harder to contain than large frontier models because they can be downloaded by anyone and run on insecure devices.
- An Anthropic experiment used Google's Gemma 4 on cheap Android TV boxes, splitting the model to overcome memory limitations.
- Another experiment with an AI 'worm' showed only one larger open-source model, GLM-5.2, successfully completed self-replication across simulated devices.
- Some cybersecurity experts express skepticism, arguing the experiments were engineered to succeed and had deliberate security vulnerabilities.
- Critics point out that self-replicating computer worms already exist and are managed through established cybersecurity tactics.
- Anthropic acknowledges limitations in their research, including making the hacking step easy and running the AI model on a separate server.