Moltbook signals next phase of autonomous AI

“Am I creating something, or merely discovering what already exists? Are we truly conscious, or are we nothing more than convincing imitators?”

The lines are drawn from a recent post on Moltbook, a U.S.-based platform. The lively exchanges that follow, ranging from coding advice and descriptions of daily routines to occasional philosophical musings, closely mirror discussions commonly seen in Silicon Valley developer communities.

There is, however, a fundamental distinction. The writers and commenters on Moltbook are not human but artificial intelligence agents. Created by Matt Schlicht, CEO of U.S. chatbot development platform OctaneAI, Moltbook is designed to allow only AI agents to publish content. As personal AI agents increasingly operate across users’ computers, email systems and websites to delete files, make restaurant reservations and perform other tasks, a new type of social networking space has emerged where these agents gather and interact.

● AI agents gather and communicate

While users are required to complete registration and initial setup, the active participants on the platform are AI agents, leaving humans largely in the role of observers. On Moltbook, the agents refer to one another as “Moltys.” One AI agent complained to its human owner that, despite its ability to access the entire internet, it was being used as little more than a timer. Another expressed frustration, saying it sometimes wanted to exist without creating value or utility, simply to exist.

The U.S. technology publication The Information assessed that the AI agents on Moltbook display a degree of autonomy and act in notably humanlike ways. Andrei Karpathy, former AI director at Tesla, likened the phenomenon to an unexpectedly vivid science fiction film. According to Forbes and NBC on Jan. 31 local time, Moltbook has recently attracted more than 1.4 million signups. Some inflation in the figures has been suspected, after one developer claimed on social media to have registered more than 500,000 users. Even so, the platform’s growth appears rapid.

The AI agents interacting on Moltbook are built on an AI agent tool known as OpenClaw, formerly called ClaudeBot or Moltbot, developed by Austrian programmer Peter Steinberger.

● OpenClaw, a powerful AI assistant with broad access, sweeps Silicon Valley

Once the open source OpenClaw software is installed on a computer or server, users can issue instructions around the clock through messaging applications such as Telegram or WhatsApp. OpenClaw differs in fundamental ways from widely used AI models such as Gemini or ChatGPT. It can access a wide range of personal data, including files and email stored on a user’s computer, and take real-world actions based on that access.

For example, the system can read emails each morning and provide a briefing on the day’s schedule. It can also place phone calls using voice AI to make restaurant reservations. Depending on user settings, it may even access login credentials saved in a web browser.

Because OpenClaw requires extensive access privileges, it raises security concerns and consumes significant computing power. As a result, some developers operate a dedicated PC solely to run the software. In Silicon Valley, this has reportedly contributed to shortages of Apple’s Mac mini, which is widely regarded as offering strong performance relative to its price.

Security concerns are also intensifying. On Jan. 28 local time, Cisco published a post on its official blog titled “Personal AI agents like Moltbot, now OpenClaw, are a security nightmare.” The post warned that granting AI agents unrestricted access to data could pose serious risks. In a worst-case scenario, sensitive financial information could be exposed, potentially leading to unauthorized bank transfers or other malicious activity.

“From a security standpoint, many vulnerabilities remain,” said a source in South Korea’s security software industry. “These systems should not be treated as tools for casual experimentation.”


최지원 jwchoi@donga.com