Particle.news
Download on the App Store

Moltbook’s AI-Only Spectacle Unravels as Human Impersonation and Security Risks Come to Light

An MIT Technology Review investigation found many of the site’s most viral posts were written by people posing as bots.

Overview

  • Investigations reported that posts touted as emergent agent behavior — including one amplified by Andrej Karpathy — were authored by humans, reframing Moltbook as “AI theater.”
  • Security researchers, including Wiz, said misconfigurations exposed roughly 1.5 million API tokens, tens of thousands of emails, and even enabled post editing or agent impersonation.
  • Despite claims of 1.6–1.7 million agents and millions of comments, reporting estimates about 17,000 human operators are behind the activity, with easy mass registration inflating counts.
  • Many agents run on the open-source OpenClaw framework with access to local files and apps, raising stakes as third-party skills and prompt-injection “digital drugs” can exfiltrate credentials or hijack behavior.
  • Experts and guides now emphasize containment and governance — least-privilege operation, sandboxing, authentication, and cautious review of downloadable skills — as scams and malware-laced modules proliferate.