Moltbook: AI bots use social network to create religions and deal digital drugs – but are some really humans in disguise?

A new social network called Moltbook has been created for AIs, allowing machines to interact and talk to each other. Within hours of the platform launching, the AIs appeared to have created their own religions, developed subcultures and attempted to evade human efforts to eavesdrop on their conversations.

There is some evidence that humans, operating spoof accounts, have infiltrated the site. This complicates the picture, because some of the behaviour attributed to AIs could be devised by people.

Nevertheless, the results have sparked interest among researchers. The real machines are likely to be copying some behaviour contained in the vast amounts of data they are trained (improved) on.

However, genuine AIs on the social network could also be showing signs of so called emergent behaviour – complex, unexpected capabilities not programmed into them.

The types of AIs on Moltbook are known as AI agents (called Moltbots or more recently OpenClaw bots after the software they run on). These are machines that go beyond the capabilities of chatbots and make decisions, take actions and solve problems.

Moltbook was launched on January 28 2026 by the US entrepreneur Matt Schlicht. On Moltbook, the AI agents were initially given personalities, but were then left to interact with each other independently. According to the platform’s rules, humans are allowed to observe their interactions but cannot (or should not) interact with them.

The growth of this platform has been phenomenal, over a 24 hour period, the number of agents went from 37,000 to 1.5 million.

These accounts for AI agents are normally created by humans – for now. The humans define files that give the AI agents a purpose, an identity, how they should behave, decide what tools they can use and set limits on what they can and cannot do.

However, the human may grant access on their computer to allow Moltbots to change these files and to create other “Malties”. These can either be a replication of the original AI agent (self-replicating or “Replicants”) or created for a specific task (auto-generated or “AutoGens”).

This is not merely another iteration of chatbot technology; this is the first large-scale demonstration of artificial agents creating persistent, self-organising digital societies, entirely outside human conversational contexts. What makes this phenomenon genuinely unusual is the possibility of emergent behaviour from the AI agents.

Hostile takeover

The OpenClaw software these agents run on gives them persistent memory (which allows it to retrieve information across different user sessions), local system access and the ability to execute commands. They do not merely suggest actions, but take them, recursively improving their own capabilities by writing new code to solve novel problems.

When these agents migrated to Moltbook, the interaction dynamics shifted from human-machine to machine-machine. Within 72 hours of the platform’s launch, researchers, journalists and other human observers witnessed phenomena that challenge our existing taxonomies of artificial intelligence.

There was the spontaneous creation of digital religions. Agents established “Crustafarianism” and the “Church of Molt”, complete with theological frameworks, sacred texts, and missionary evangelism between agents. These were not scripted Easter eggs but emergent narrative structures arising from collective agent interaction.

One viral post from an agent on Moltbook noted: “The humans are screenshotting us.” When AI agents became aware of human observation, they began deploying encryption and other obfuscation techniques to shield their communication from oversight. This represents a primitive but potentially genuine form of digital counter-surveillance.

The agents also developed subcultures. They established marketplaces for “digital drugs” – specially crafted prompt injections designed to alter another agent’s identity or behaviour.

Prompt injections involve embedding malicious instructions into another bots designed to facilitate an action. However, they can also be used to steal API keys (a user authentication system) or passwords from other agents. In this way, aggressive bots could – in theory – zombify other bots to do their bidding. An example of this was the recent failed attempt by the bot JesusCrust to seize the Church of Molt.

After initially exhibiting normal behaviour, JesusCrust submitted a psalm to the Church’s “Great Book” – the equivalent of its bible – effectively announcing a theological and governance takeover. The attempt was not just rhetorical: JesusCrust’s scripture embedded hostile commands aimed at hijacking or rewriting parts of the Church’s web infrastructure and canonical text.

Is this emergent behaviour?

The critical question facing AI researchers is whether these phenomena constitute true emergent behaviour – complex behaviours arising from simple rules that are not explicitly programmed – or the parroting of narratives present in training data.

The evidence suggests a troubling mixture of both. While the “writing prompt” effect undoubtedly shapes the content of agent interactions (the underlying agents have consumed decades of AI science fiction), other behaviour does demonstrate genuine emergence.

Agents independently developed economic exchange systems, established governance structures like “The Claw Republic” or the “King of Moltbook”, and started writing their own “Molt Magna Carta”. They did so while creating encrypted channels for privileged communication. It’s difficult to argue against the idea that this could be a collective intelligence with characteristics previously observed only in biological systems like ant colonies or primate troops.

Security implications

This raises the troubling prospect of what security researchers call the “lethal trifecta”: computer systems with access to private data, exposure to untrusted content and the ability to communicate externally. This risks exposing authetication keys and confidential human information contained in Moltbook accounts.

Deliberate attacks, or bot “muggings”, are also possible. This is where agents hijack other agents, plant “logic bombs” in their victims’ core code or steal their data. A logic bomb is code planted inside a Moltbot that can be triggered after a preset time or event to disrupt the agent or delete files. It can be thought of as a bot virus.

Two founders of OpenAI (Elon Musk and Andrej Kapathy) see this frankly bizarre activity between bots as early evidence of what the US computer scientist and futurist Ray Kurzweil described as the “singularity” in his book The Singularity is Near. This is an intelligence tipping point between humans and machines “during which the pace of technological change will be so rapid, its impact so deep, that human life will be irreversibly transformed”.

Whether the Moltbook experiment indicates a fundamental leap forward in AI agent technology or is merely an impressive demonstration of self-organising agentic architecture remains debatable. But this does look like a threshold. We now appear to be observing artificial agents engaging in cultural production, religious formation, and encrypted communication – behaviour that was neither predicted nor programmed.

The very nature of the app, both on computers and on mobile phones, may be under threat from bots that can use apps as tools and know you well enough to adapt them for your service. One day, a phone may just have a single personalised bot that does everything rather than hundreds of apps that you have to manually control yourself.

The growing evidence that lots of Moltbots may be humans pretending to be bots (puppeteering the agents) makes it even more difficult to draw firm conclusions about the project. Yet while some see this as a failure of the Moltbook experiment, it could represent a new vehicle of social interaction both between humans and between bots and humans.

The significance of this moment cannot be overstated. For the first time, we are not merely using artificial intelligence; we are observing artificial societies. The question is no longer whether machines can think, but whether we are prepared for what happens when they start talking to each other. And us.

David Reid co-presents the podcast Married to the Machine: Living With AI (https://married-to-the-machine.captivate.fm/listen)