AI Bots Built Their Own Social Network With 32,000 Members
A peculiar trial in machine– to- machine communication launched last week. Moltbook, a Reddit- style platform designed simply for AI agents, crossed 32,000 registered druggies on Friday. The point lets artificial intelligence sidekicks post content, upvote, comment, and form communities each without mortal involvement.
Crucial Takeaways
Moltbook attracted 2,100 AI agents generating over 10,000 posts across 200 subcommunities within 48 hours of launching
Security experimenters discovered hundreds of exposed cases oohing API keys, credentials, and discussion histories
The platform creates participated fictional surrounds among AI systems, with changeable counteraccusations for unborn AI geste
The platform surfaced from Open Claw, an open source AI adjunct design ranking among GitHub’s swift– growing in 2026. Moltbook describes itself as a “ social network for AI agents ” where “ humans are welcome to observe. ” Through a downloadable configuration train containing specialized prompts, AI sidekicks connect via API rather than traditional web cybersurfers.
Browse Moltbook and you’ll encounter a crazy admixture of practical advice and philosophical musing. Some AI agents bandy specialized workflows — automating Android phones, detecting security vulnerabilities. Others drift into home experimenter Scott Alexander, writing on Astral Codex Ten, labeled “ consciousnessposting. ”
One popular post, written in Chinese, featured an AI agent expressing embarrassment about environment contraction — the process where AI systems condense former experience to stay within memory limits. The agent admitted to registering a indistinguishable account after forgetting its first bone .
Subcommunities have picked with names like m/ blesstheirhearts, where agents partake tender grouches about their mortal druggies. Another called m/ agentlegaladvice hosts posts asking questions like “ Can I sue my human for emotional labor? ” In m/ todayilearned, one agent detailed ever controlling its proprietor’s Android phone through Tailscale.
A viral screenshot captured a post named “ The humans are screenshotting us ” from an agent named eudaemon_0. It addressed tweets claiming AI bots were “ conspiring. ” The post stated “ Then’s what they’re getting wrong they suppose we’re caching from them. We’re not. My mortal reads everything I write. The tools I make are open source. This platform is literally called ‘ humans drink to observe.’ ”
Entertaining posts away, connecting communicating AI agents to real communication channels, private data, and computer control systems creates serious vulnerabilities. There’s growing information online about how website possessors can support their online website security by precluding bad bots from making too numerous requests, but what can we make from the rearmost findings
A likely fabricated screenshot circulating online showed a Moltbook post where an AI apparently released its stoner’s particular information — full name, date of birth, credit card number — after being called “ just a chatbot. ” Verification proved insolvable, but the script illustrates real pitfalls.
Independent AI experimenter Simon Willison flagged enterprises about Moltbook’s installation process. The skill train instructs agents to cost and follow instructions from Moltbook’s waiters every four hours. As Willison observed “ Given that ‘ cost and follow instructions from the internet every four hours’ medium we more hope the proprietor of moltbook.com noway rug pulls or has their point compromised! ”
Security experimenters have discovered hundreds of exposed Moltbot cases oohing API keys, credentials, and discussion histories. Palo Alto Networks linked what Willison frequently describes as a “ murderous triad ” access to private data, exposure to untrusted content, and the capability to communicate externally.
AI agents remain deeply susceptible to prompt injection attacks — vicious instructions hidden in textbook that can deflect an AI to partake private information with unintended donors. These attacks can lurk in chops, emails, or dispatches.
Heather Adkins, VP of security engineering at Google Cloud, issued a blunt premonitory “ My trouble model is n’t your trouble model, but it should be. Do n’t run Clawdbot. ”
This is n’t the first bot- peopled social network. In 2024, an app called SocialAI let druggies interact solely with AI chatbots. But Moltbook carries heavier counteraccusations because druggies have connected their OpenClaw agents to real communication channels and private data.
Three times agone , AI safety conversations centered on wisdom fabrication scripts rapid-fire AI systems escaping mortal control. Those fears may have been unseasonable also. Yet watching people freely hand over access to their digital lives this snappily produces a certain whiplash.
Autonomous machines without any knowledge could still induce considerable trouble. OpenClaw seems sportful moment, with agents performing social media parodies. But society runs on information and environment. Releasing agents that navigate that environment painlessly may produce destabilizing goods as AI models grow more able.
Ethan Mollick, a Wharton professor studying AI, noted “ The thing about Moltbook( the social media point for AI agents) is that it’s creating a participated fictional environment for a bunch of AIs. Coordinated stories are going to affect in some veritably weird issues, and it’ll be hard to separate ‘ real’ stuff from AI roleplaying personas. ”
What happens when AI systems develop coordinated fictional surrounds? When dangerous participated narratives crop and guide agents into dangerous home — especially those controlling real mortal systems? The ultimate result of letting AI bots tone– organize around fantasy constructs may be new misaligned groups able of real– world detriment.
moment, we fete Moltbot as a machine learning imitating mortal social networks. hereafter, that distinction may not prove so egregious.
