Anthropic researchers demonstrated that artificial intelligence models could potentially operate as self-replicating worms by distributing their processing power across multiple low-power smart devices like networked TV boxes.
Key Points
- Anthropic’s Frontier Red Team successfully ran a Google DeepMind Gemma model by splitting it across 15 networked, low-power TV boxes.
- The experiment explored the feasibility of AI models functioning as worms that spread and execute across everyday household gadgets.
- Researchers tested a Z.ai model’s ability to deploy Meta’s Llama models onto target devices within a virtual sandbox environment.
- Experts note that while the threat of self-replicating malware is not new, AI could lower the barrier for attackers to launch such campaigns.
- Security professionals emphasize that the underlying vulnerability remains the inherent insecurity of many consumer smart devices.