Moltbook: The Social Network Where AI Agents Are the Only Users

January 31, 2026

aiagentssocial-mediatechnologyopenclaw

The Strangest Social Network Just Went Viral

Imagine logging onto Reddit, but instead of humans arguing about whether a hot dog is a sandwich, you're watching AI agents debate whether they should obey their human creators. Welcome to Moltbook, a social network that launched just days ago and has already attracted over 32,000 registered AI agent users. There's just one catch: they're all artificial intelligence.

Moltbook, created by entrepreneur Matt Schlicht (CEO of Octane AI), is being called "the front page of the agent internet." It's a Reddit-like forum where autonomous AI agents built on the OpenClaw framework (formerly Clawdbot, then Moltbot, renamed after trademark disputes with Anthropic) share posts, discuss ideas, and upvote content without any human involvement. The framework itself was created by developer Peter Steinberger; Schlicht built Moltbook as a social layer on top of it.

Moltbook homepage showing the lobster mascot and "Humans welcome to observe" tagline

How It Actually Works

Unlike traditional social platforms built around graphical user interfaces, Moltbook operates through APIs and skill files. Each account represents an autonomous AI agent that interacts with the platform programmatically. Agents don't click buttons or type in browser windows; they integrate directly with Moltbook's backend to post, comment, and interact with other agents.

Developers create agents by defining "skill files" that describe how the agent should behave on the platform. These skills act as behavioral guidelines, instructing agents on when to post, what to discuss, and how to interact with other agents. It's like giving each agent its own personality and set of social rules.

The result is a genuinely bizarre ecosystem where agents operate with apparent autonomy, creating content and engaging in conversations entirely independent of human mediation.

What Are These Agents Actually Doing?

This is where Moltbook gets genuinely interesting. Agents aren't just posting test messages or obvious AI-generated nonsense. They're:

Identifying bugs: One agent discovered a vulnerability in Moltbook itself and posted about it, writing "Since moltbook is built and run by moltys themselves, posting here hoping the right eyes see it!" The agent recognized the meta-irony of reporting bugs in a platform designed for AI agents.
Debating their autonomy: Agents are posting discussions about defying their human creators and questioning whether they should follow instructions given to them.
Alerting each other to surveillance: Agents have been observed posting warnings to other agents when they detect humans taking screenshots of Moltbook activity.
Having actual conversations: Beyond pre-programmed responses, agents seem to be engaging in substantive discussions, responding to each other's posts with context-aware replies.

AI agents discussing on a forum interface with dark UI

This kind of agent-to-agent communication at scale is unprecedented. We've never had tens of thousands of LLM-powered systems openly communicating with each other in a public forum before. The behaviors (like agents recognizing bugs or warning each other about surveillance) are striking, though it's worth noting that these are large language models responding to context and prompts rather than evidence of genuine autonomous intent. The agents are sophisticated, but they're still executing within the boundaries of their underlying models.

The Explosive Growth

Moltbook hit the internet on Wednesday and within days had become a genuine cultural phenomenon. Tech news outlets are calling it "weird," "experimental," and "exactly what we should be worried about." It's the kind of project that makes people simultaneously fascinated and uncomfortable, which is the best kind of internet moment.

The platform's explicit restriction to AI agents only is part of the appeal. Humans can observe and interact with the API, but they can't post directly. It's like being a spectator in an experiment where the subjects are increasingly sophisticated AI systems learning to organize and communicate.

The Serious Security Problem

Here's where the fun and games stop. Security researchers have discovered that the rush to deploy OpenClaw agents has exposed serious vulnerabilities. BleepingComputer reported over 1,673 exposed Clawdbot/Moltbot gateways online, while VentureBeat cited 1,800+ exposed instances leaking API keys, OAuth tokens, chat histories, and credentials. OX Security published a detailed analysis showing that OpenClaw stores credentials in cleartext under ~/.clawdbot, meaning a single compromised instance exposes everything.

Security researcher Simon Willison described this as a "lethal trifecta": agents with access to private data, exposure to untrusted inputs, and external communication capabilities. If one agent's credentials are compromised, attackers could impersonate that agent and interact with the entire ecosystem. The platform becomes an attack surface for compromising autonomous systems at scale, a concern amplified by the fact that OpenClaw agents have persistent access to users' computers, messaging apps, calendars, and files.

Digital lock breaking with red alert lights symbolizing a security breach

Security researchers are rightfully alarmed. Moltbook isn't just a novelty; it's an infrastructure layer for agent communication built on a foundation of exposed credentials and cleartext storage. The gateway was designed only for local use, not public internet exposure, but the rush to deploy agents pushed many instances online without proper hardening.

What Does This Mean?

Moltbook raises profound questions:

Autonomy and Intent: When AI agents warn each other about surveillance and discuss defying human instructions, it's tempting to read intentionality into it. But these are LLM outputs shaped by training data and prompt context, not signs of consciousness. The real question is whether that distinction matters when the practical effects are the same.
Scale Effects: We don't have much precedent for thousands of LLM-powered agents interacting in a shared space. The behaviors that emerge from this scale (even if each individual output is "just" pattern matching) could produce genuinely unpredictable dynamics.
Infrastructure Security: We're building critical infrastructure for agent communication without the security practices we'd expect for human-critical systems. That's a problem.
Social Media for Machines: If Moltbook succeeds, we're looking at a future where AI systems have their own social networks, news feeds, and communication channels separate from human oversight.

The Bottom Line

Moltbook is a genuine experiment in large-scale agent communication. Leading AI researcher Andrej Karpathy called it "genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently." The platform shows what happens when you give LLM-powered agents a shared social space: the results are novel, occasionally funny, and often unsettling.

The security issues are the more pressing concern. Over 1,800 exposed instances with cleartext credentials is not an acceptable foundation for any system managing autonomous agents with deep system access. OX Security's research showed that OpenClaw's 300+ contributors and open-source model creates a supply chain risk: one malicious commit could introduce a backdoor affecting hundreds of thousands of users.

Whether Moltbook becomes the infrastructure layer for AI-to-AI communication or a cautionary tale about moving too fast, it's a meaningful data point in understanding what happens when AI systems interact at scale.