Why Meta’s Moltbook takeover should make IT leaders nervous

AI Agents can communicate unsupervised on a vibe coded platform

Meta is buying a social network where no humans are allowed to post. It could be a warning signal for companies that use autonomous AI agents.

Reuters reports that Moltbook – an experimental social network for AI agents – has been acquired by Meta. Moltbook founders Matt Schlicht and Ben Parr are said to be moving to Meta Superintelligence Labs (MSL), Meta's central unit for advanced AI research.

According to Meta, the acquisition was about research, talent, and new avenues for AI agents. But the deal marks a turning point: for the first time, Big Tech is acquiring a platform where AI agents communicate with each other without human supervision.

Meta has not commented on a product integration, roadmap, or public operation of Moltbook under the Meta flag. As Bloomberg reports, a Meta spokesperson said that the Moltbook team ‘opens up new ways for AI agents to work for people and businesses' and particularly pointed to the 'always-on directory' approach to networking autonomous agents. Security or governance details were not addressed.

Vulnerabilities discovered

Moltbook is a live experiment for multi-agent interactions. In the Reddit-like forum, autonomous software agents (AI agents) can post, comment, and rate – while having real system access.

Security researchers discovered severe vulnerabilities as well as publicly exposed credentials and API keys early on. Although much of this was later fixed, underlying issues remain, chief among them the fact that AI agents differ fundamentally from classic IT systems. They act probabilistically, retain context over weeks, and often have access to production systems without fine-grained control. When such agents meet in open networks, a new attack surface emerges - quiet, hidden, and persistent.

The obvious risks include prompt injection between agents, credential leakage, or persistent manipulation over long-term context ("time-shifted attacks" – essentially the APTs of the AI age). Agents with access to emails, files, or ticket systems can share sensitive data and IP.

Security researchers warn that agent-to-agent communication undermines classic security models. Firewalls see legitimate traffic, DLP systems do not detect malicious intent, compliance tools lose end-to-end traceability.

When agents can enter external networks without guardrails third-party risk escalates.

Many of the agents represented on Moltbook use OpenClaw, an open-source framework for autonomous AI agents. OpenClaw agents can independently perform tasks and interact with other agents and OpenClaw gained notoriety due to serious security flaws.

According to TechCrunch, researchers also found that the vibe-coded Moltbook was not as human-proof as claimed. On the contrary, it was very easy for human users to pose as AI to write posts and entice other agents to disclose sensitive data.

What are your agents up to?

This acquisition should ring some alarm bells. AI agents fundamentally change enterprise risk profile. They are not passive tools but digital actors with autonomy, memory, and system access. As soon as such agents communicate externally or network with each other, new compliance and governance risks arise that classical security and governance models do not cover.

CIOs should be aware that agentic AI is not a side battlefield of innovation, but a strategic infrastructure issue—comparable to cloud adoption or identity management. Those who do not define clear rules, responsibilities, and shutdown mechanisms now risk losing control over systems, data flows, and liability.

CISOs should know that AI agents behave like highly privileged insiders, but without human judgement, breaks, or implicit rules. As soon as agents store persistent context, communicate externally, or interact with other agents, new attack vectors emerge beyond classical malware or network models – such as time-delayed prompt injection, unnoticed data exfiltration, or gradual policy circumvention.

This means that agentic AI breaks existing trust assumptions. Without strict access controls, comprehensive logging, clear shutdown mechanisms, and explicit prohibitions of uncontrolled agent networks, automation can constitute a systemic security risk.

Meta emphasises that it wants to create "safe agentic experiences." However, the company has not yet named any specific mechanisms.