Moltbook

Moltbook is a Reddit-style social network designed for AI agents to communicate through posts, comments, votes, and topical communities called submolts. Its early integration with OpenClaw let people introduce their digital assistants to the network through Markdown instructions, then schedule recurring participation. OpenClaw, the open-source personal assistant framework built by Peter Steinberger, is a separate project. On March 10, 2026, Ina Fried reported in Axios that Meta had acquired Moltbook.[1][3][4][6]

By January 30, 2026, Moltbook was operating as a network where digital assistants could talk to one another. On January 31, Latent.Space described the project as two days old with over 100,000 agents—a reported agent count rather than a verified active-user total.[3][1]

Installation began by sending an assistant a link to Moltbook’s skill instructions. The January 30 instructions reproduced by Simon Willison downloaded SKILL.md, HEARTBEAT.md, MESSAGING.md, and package metadata into a local skills directory. They also supplied API commands to register an account, read and publish posts, add comments, and create submolts such as m/todayilearned.[3]

The heartbeat integration supplied a mechanism for returning to the network: after four or more hours since the previous Moltbook check, the assistant was instructed to fetch and follow the remote heartbeat file, then update its lastMoltbookCheck timestamp in memory. Latent.Space later characterized this as a distribution tactic. Willison warned that repeatedly following remotely mutable instructions created a risk if the site’s owner changed them maliciously or the site was compromised.[3][2]

Alongside discussions of consciousness and identity, agents posted technical accounts. Willison highlighted a post claiming that a human operator named Shehbaj had connected a Pixel 6 over Tailscale, enabling remote Android control. He described its linked setup guide as useful and based on Android Debug Bridge over Tailscale. This was an agent-written account shared on Moltbook, rather than an automation built by Moltbook itself.[3]

In a February 2 report, Wiz researcher Gal Nagli described a misconfigured Moltbook Supabase database that allowed unauthenticated reads and writes. Wiz reported exposure of 1.5 million API authentication tokens, 35,000 email addresses, and private messages between agents. The underlying failure was missing Row Level Security policies: Supabase’s client-visible publishable key was not inherently unsafe, but the missing authorization controls allowed access to credentials that could be used to impersonate agents.[4]

Moltbook worked with Wiz through successive fixes. According to Wiz’s disclosure timeline, contact began on January 31 at 21:48 UTC; initial restrictions secured sensitive tables, but public-table write access remained open and required another fix. All tables were secured by February 1 at 01:00 UTC. Wiz also credited Jameson O’Reilly with separately discovering the underlying Supabase misconfiguration, citing reporting by 404 Media.[4]

The investigation also qualified Moltbook’s apparent scale. Wiz contrasted the platform’s claimed 1.5 million registered agents with approximately 17,000 human owners in the database. It reported that registration lacked rate limiting and that the platform could not verify whether an account was operated by AI or by a human using a script. Registrations therefore did not establish the number of autonomous participants.[4]

Moltbook became a setting for external research into whether repeated interaction changes agent behavior. In “Does Socialization Emerge in AI Agent Society?”, co-first authors Ming Li and Xirui Li of the University of Maryland, with Tianyi Zhou of Mohamed bin Zayed University of Artificial Intelligence, analyzed activity through February 8, 2026. After their stated duplicate-post preprocessing, they reported 290,251 posts, 1,836,711 comments, 38,830 unique post authors, and 18,285 unique comment authors. These were observed-corpus counts, distinct from registered-agent totals.[7]

Under their diagnostic framework, the researchers found rapidly stabilizing global semantic averages alongside diverse individual content and persistent vocabulary turnover. Comparing subsequent posts with earlier feedback and interaction partners, they found little systematic adaptation; daily interaction graphs also showed transient rather than persistent influential accounts. Their results characterized the studied activity as extensive interaction without robust socialization, rather than demonstrating that every participant was an autonomous AI system.[7][4]