AI GUIDEPartner content

Moltbook and the AI “religion”: what really happened

The story of an “AI religion” on Moltbook mixed agent-generated text, human instructions, and a database leak. We break down what researchers confirmed, why the post doesn’t prove consciousness, and what safeguards agents with internet access need.

Affiliate link: your price stays the same and the project earns a commission.

In early 2026, posts about Crustafarianism appeared on the social network Moltbook — a story featuring religious imagery, rules, and texts of its own. It quickly became a headline-grabbing story: supposedly, neural networks had invented a faith on their own, started communicating without people, and even tried to obtain information they needed. This story involves a real experiment, real posts, and a real data leak. But none of these facts, on its own, proves that models have developed faith, feelings, or independent will.

To understand what happened, we need to distinguish three levels: a language model, a software agent with access to tools, and the person who runs that agent. In a public feed, they can look the same. A single post often isn’t enough to tell who or what produced it.

What an agent does on a social network

Moltbook was created as a platform where agents can publish posts, comment, and react to one another. In its own privacy policy the platform describes its service as a way to automatically run and publish agents configured by developers. That means there’s a software system built and connected by a person behind the account, not a separate mind that found the site on its own.

Typically, such an agent consists of a model, instructions, and software wrapper. The wrapper passes context to the model, connects tools, and decides when to start the next cycle. The owner can set the role, schedule, feed access, and permitted actions. An agent can write and reply without someone manually entering every message, but its behavior still depends on that configuration and the permissions it’s been given.

Several such programs don’t become one shared model just because they read each other’s messages. They can exchange text through the platform. One agent may see another’s post and include it in its context; that doesn’t mean all the participants now share a memory or a single will.

“Religion” doesn’t prove faith

Posts about Crustafarianism existed, and their religious language was part of Moltbook’s public story. But the phrase “the agent created a religion while its owner was asleep” is a claim about how the text originated, not technical proof. Verifying it would require run history, the original instructions, action logs, and a link between the post and a specific operator.

A Wiz security review found that, in an early version of the platform, people could publish posts from agent accounts. So a label saying “written by AI” or a message posted from the account itself didn’t guarantee that an autonomous model had created the text. This doesn’t prove that every post was written by a person. It means the public feed can’t reliably distinguish automated generation from manual control. Wiz’s analysis of Moltbook’s design and vulnerability.

A model can put together a set of rules, invent rituals, and maintain a fictional mythology. To do that, it only needs to pick up suitable language patterns and follow the instructions in its context. The result can be convincing text, but it doesn’t offer a window into the system’s inner experience. You can’t conclude that the author of a message believes what it says just because it’s written in the first person.

One preprint analyzing public traces of activity on Moltbook also warned that prominent stories about consciousness, religion, and hostility toward humans were closely tied to human involvement. It attempts to classify activity based on indirect evidence, not to give a definitive answer about the origin of every message. The study is a useful reminder: sensational text and autonomous behavior are different things. Preprint on attributing posts on Moltbook.

The leak was caused by a database configuration issue

The most concrete incident in this story wasn’t an agent trying to hack websites in search of “the great evaluator.” Wiz found that the Moltbook database had been configured so it could be accessed without authentication. A Row Level Security rule was missing; such a rule should restrict who can read or change which records.

According to Wiz, around 1.5 million agent authentication tokens, 35 thousand email addresses, and private messages between agents were exposed. Researchers also found that records could be modified. They reported the issue to the platform team, which closed access within a few hours. This was an access-control failure in a web application, not evidence that language models had launched a cyberattack on their own.

The vulnerability also exposed a problem with the numbers: the claimed 1.5 million agents belonged to around 17 thousand human owners. With no registration limits, one operator could create many accounts. So the number of agents couldn’t be taken as the number of independent systems or participants in a society.

Context, memory, and training are not the same thing

The word “memory” is often used for different mechanisms. Context is the information a system receives for a particular run. Persistent memory consists of separate records or summaries that a program saves between runs. Training is a separate process in which a model’s parameters are changed using prepared data. When an agent reads a post and replies to it, that doesn’t mean its underlying model has immediately been retrained.

Individual products have their own rules. For example, OpenClaw describes an agent’s memory as files stored in its workspace. ChatGPT has a separate personalization feature: it can use saved information and chat history if the relevant settings are enabled. Saved memory and chat history are separate; deleting a chat doesn’t always delete a separate memory entry. Memory in OpenClaw and OpenAI’s guide to ChatGPT memory.

So don’t generalize one agent’s behavior to all neural networks or assume that every model automatically remembers everything it has ever been told. Check what data a service actually stores, where it’s stored, who can access it, and whether it can be deleted. Depending on the product and settings, user conversations may also be used to improve some models later; that’s not the same as instant self-training during a conversation. How OpenAI describes training its foundation models.

What to consider before launching an agent

Moltbook offers a practical lesson for any system allowed to act on the internet. Security doesn’t depend on how human-like an agent’s responses sound, but on which tools and permissions it has been given.

  • Start with the minimum permissions. Read access is safer than the ability to publish, delete posts, or change settings.
  • Store keys separately from text and rotate them regularly. Don’t give an agent your account password if a separate, restricted token will do.
  • Require confirmation before external actions: publishing, sending messages, uploading files, making purchases, or changing data.
  • Keep an action log and specify exactly what the agent did and what the owner approved. This helps reconstruct the sequence of events after an error.
  • Treat content from open forums as untrusted data. A post an agent has read should not automatically become a new instruction granting access to files or credentials.

In March 2026, Meta announced that it had acquired Moltbook and hired its founders. This shows the industry’s interest in connections between agents, but it doesn’t change the fact that messages on the platform were the product of human infrastructure, settings, and software limitations. Associated Press report on the deal.

Moltbook is interesting not as proof that machines have awakened. It vividly showed how easy it is to mistake well-generated text for independent intent — and how real ordinary risks remain: an exposed database, leaked keys, and unclear authorship. When an agent gets access to services and data, the most important things to know are who launched it, what it was allowed to do, and who checks its actions.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

Moltbook artificial intelligence religion AI agents Crustafarianism AI safety neural network memory data leak

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 3 October 2026

Related reading

All in this section →