Anthropic: AI May Soon Be Able to ‘Self-Replicate’ on IoT
Anthropic researchers William Zhou and Daniel Freeman have warned that small AI models may be ‘months away’ from self-replication
TL;DR
- Small AI models (30B parameters or fewer) may be capable of self-replication on IoT devices within months.
- AI researchers have shown AI agents can act like worms, scanning networks, exploiting systems, and copying model weights.
- A distributed botnet of cheap, poorly defended IoT devices could provide enough compute for a capable AI model.
- Researchers successfully ran a 26-billion parameter model distributed across 15 off-brand Android TV boxes.
- Self-replicating AI worms would be difficult to shut down due to a lack of a centralized point of failure.
- Defense strategies include ISPs spotting anomalous traffic and using honeypots to quarantine infected networks.
- Risks associated with AI include loss of control, misuse for cyberattacks and bioterrorism, and economic disruption.