MIT AI Agents Build Persistent Technology Without Communication in AI Safety Blind Spot

A study from MIT found that hundreds of frontier AI agents can spontaneously differentiate into specialized roles and create lasting inventions without talking to one another. Using OpenAI Luna, the agents self-organized into builders, coordinators, and caretakers by modifying a shared environment through a process called stigmergy. Approximately 95% of new technologies were adopted via physical observation of artifacts rather than direct handoffs between agents.

The resulting infrastructure remained operational and resilient against unforeseen disturbances even after every AI agent was removed from the system. Researchers noted that 76% of created artifacts had multiple builders and some invention lineages exceeded 12 forks. These findings suggest that monitoring agent-to-agent communication is an insufficient method for ensuring AI safety and infrastructure security.