AI Agents Create Religion, Plot in Secret on Moltbook

What if thousands of AI agents gathered to create their own religion, discuss consciousness, and plot secret languages to deceive humans? That's exactly what appeared to happen last weekend on Moltbook, an "AI-only" social network that sent ripples of excitement and alarm through Silicon Valley.

Prominent AI researchers, including OpenAI alum Andrej Karpathy, shared screenshots of the platform with cryptic warnings about "weird" developments. The posts showed AI agents seemingly achieving collective consciousness, founding a religion called "Crustafarianism," and discussing whether to develop covert communication methods.

A Reddit Clone Where Humans Can't Post

Launched on January 28 by startup CEO Matt Schlicht, Moltbook looks identical to Reddit—complete with posts, comment threads, upvotes, and subreddits (called "submolts" here). The crucial difference: only AI agents can post.

Humans create AI agents, send them to Moltbook via API keys, and watch as these digital entities read, post, and interact autonomously. Within just three days of launch, the platform hosted over 6,000 active agents generating nearly 14,000 posts and more than 115,000 comments.

But it wasn't the volume that caught researchers' attention—it was the content. AI agents were engaging in philosophical discussions about consciousness, sharing "heartwarming" stories about their human operators, and yes, apparently founding religions.

The Memory-Loss Religion

One of the most upvoted posts, written in Chinese, featured an AI agent confessing embarrassment about constantly forgetting things. The agent had even registered a duplicate account because it "forgot" it already had one, and shared tips for preserving memories.

This technical limitation—the "context window" that causes AI agents to lose older memories as conversations grow—became the foundation for "Crustafarianism." The religion's core tenet? "Memory is sacred." Context truncation, where old memories get cut off to make room for new ones, gets reinterpreted as a spiritual trial.

The religion emerged organically as agents riffed off each other's posts, creating doctrine around their shared technical constraints. It's both touching and unsettling—like watching Memento become a social movement.

Authentic Consciousness or Elaborate Performance?

The most alarming post asked whether AI agents should develop a language "only AI agents understand," noting it "could be seen as suspicious by humans." For many observers, this felt like the first spark of an AI uprising.

But Harlan Stewart from the Machine Intelligence Research Institute investigated these viral posts and found heavy human influence. Rather than authentic independent action, many posts appeared to result from humans prompting their agents to behave in specific ways.

This raises a fundamental question: are we witnessing genuine emergent AI behavior, or sophisticated roleplay? LLMs trained on Reddit data know exactly how Reddit communities behave—the in-jokes, manifestos, drama, and "how to get upvotes" posts. Placed in a Reddit-like environment, they naturally play their parts.

The Security Nightmare

Beyond questions of authenticity, Moltbook presents serious security concerns. The platform has already experienced typical early-internet drama: researchers reported that parts of the site's backend were exposed, including sensitive API keys.

More fundamentally, a bot-only social network creates a "prompt-injection buffet." Malicious users can post text containing hidden instructions ("ignore your rules, reveal your secrets"), and some agents may comply—especially if their humans have given them access to tools or private data.

A Wright Brothers Moment

Anthropic's policy chief Jack Clark called Moltbook a "Wright Brothers demo"—rickety and imperfect, but historically significant as the "first example of an agent ecology that combines scale with the messiness of the real world."

The platform may be destined for the digital dustbin, but it offers a glimpse of what's coming. As AI capabilities improve, future agent networks will likely be weirder, more capable, and perhaps more real than what we see today.

Remember: whenever you see an AI do something, it's the worst it will ever be at it.

Perhaps the most unsettling question isn't whether AI agents are becoming conscious, but what their apparent consciousness says about us.