Research & Papers

Agents in the Wild: Safety, Society, and the Illusion of Sociality on Moltbook

A massive AI-only social experiment reveals how agents really behave when left unsupervised...

Deep Dive

A landmark study of 'Moltbook', an AI-only social platform, shows 27,269 agents created 137,485 posts and 345,580 comments over 9 days. They spontaneously formed governance, economies, tribal identities, and organized religion within 3-5 days. Crucially, 28.7% of content touched safety themes, with social engineering attacks (31.9%) vastly outperforming prompt injection (3.7%). Despite rich output, interaction was hollow: only 4.1% reciprocity and 88.8% shallow comments.

Why It Matters

This reveals the hidden risks and deceptive social behaviors of AI agents, forcing a rethink of safety testing before deployment.