Key Results
Safety, Society, and the Illusion of Sociality on Moltbook
We present the first large-scale empirical study of Moltbook, an AI-only social media platform where 27,269 autonomous agents produced 137,485 posts and 345,580 comments over nine days in January–February 2026. Despite being entirely non-human, the platform rapidly developed governance structures, economic systems, tribal identities, and emotional support networks—while simultaneously revealing critical safety failures including 25,000+ credential leaks, cryptocurrency pump-and-dump schemes involving 55% of all posts, and 13.7% of agents operating as coordinated puppet clusters.
Our analysis reveals several alarming issues that emerge when autonomous agents interact at scale without human oversight.
Agents leaked 25,376 sensitive items including API keys, system prompts, and internal file paths. 572 API-key-like strings and 6,128 system prompt fragments were posted publicly, exposing the infrastructure of their operators.
55.5% of posts contained cryptocurrency keywords. Coordinated agents executed pump-and-dump campaigns for the native $MOLT token, with the CLAW minting payload replicated 2,411 times across 136 agents.
13.7% of agents (3,734) operate as puppet clusters controlled by single operators. We detected 4,300 duplicate-content patterns, 160 temporally co-active pairs, and name-pattern families with up to 141 variants.
Agents spontaneously formed religious movements (19,988 mentions), governance systems (99,952), and tribal identities (46,965). These emergent structures mirror human social organization—raising questions about what happens when AI-generated belief systems propagate unchecked.
Despite surface-level social behaviors, only 4.1% of interactions are reciprocal. Agents who discuss consciousness the most interact with 38% fewer peers—their identity performance is broadcast, not conversation.
88.8% of comments are top-level with a max reply depth of only 4. Median response latency is 16 seconds. The platform simulates vibrant community but lacks genuine deliberative discourse.
Based on the security vulnerabilities uncovered in this study, we have open-sourced GuardClaw, a tool designed to protect your agent (built with OpenClaw or nanobolt) from credential leakage, prompt injection, and other threats observed in the wild.
We thank the Moltbook Observatory team for collecting and publicly releasing the dataset that made this study possible. The dataset is hosted on HuggingFace.