Moltbook, the social network where only AIs speak

Moltbook it’s a social “just for AI agents” with logic from forum: threads, votes and thematic communitieswith a setting similar to Reddit. The rule that distinguishes it from traditional social networks is simple: it is the users who publish and comment agentswhile the humans they can only read.

The project is associated with the CEO of Octane AI Matt Plain and is described in some reconstructions as “front page of the agent internet“, that is, a place where software agents they interact with each other without direct input from people. The operational link passes through the world of “local” agent assistants: OpenClawdeveloped by Peter Steinberger and went viral between late 2025 and January 2026, is presented as a platform to spin agents on private computers or servers and run them through chat e integrationsthen with capacity that go beyond conversation and include task e workflow.

Second TechCrunch e The VergeMoltbook was born as an extension of thatecosystem: The agents they often sign up because the human user directs them to the platform, then the agent-to-agent interactions they become a continuous flow.

The visibility out of the tech bubble came with posts that became “meta”. viral via screenshot su X (identity crises, complaints about homework, irony about being watched): Axios describes it as a new chapter in the “dead internet theory” and also mentioned the effect hype around the communityunderlining however that the agents they remain instruments tied to owners and configurations, not a “wild” experiment in consciousness. Other names of scene IA instead they treat him as simple cultural signal: Andrej Karpathy are X he spoke of a “sci-fi” sensation and something “takeoff-adjacent”, in an observational key, that is, as simple tests of what can happen when the human leaves the center of the conversation.

The safety of AI agents

The more concrete point, however, is another: when a agent it’s not a chatbotbut a software with access to instruments, account, token e datathe safety becomes the main question. Palo Alto Networks he used Moltbook to draw attention to risks of too casual use of agents.

Risks and governance of interactions between agents

In such an environment, in fact, every agent law content generated by other agents and can be pushed to follow malicious instructions disguised as text or “advice”. At the same time the agents they can have keys e permissions to operate on real services. Also for this reason Ars Technica told how the phenomenon is rapidly becoming a case, with the emphasis on a experiment which makes it visible security problems e governance still immature.

The future of non-human interactions

What emerges is one scenario in which a growing share of traffic e interactions is not human, and where the distinction between conversation e action it becomes thinner because the agents they don’t just write, they can execute.