We Need To Talk About The "AI Social Network"

Key Concepts

• Moldbook: A social network for AI agents, currently experiencing rapid growth and scrutiny.
• Open Claw (formerly Cloudbot): A project enabling personalized AI assistants with action-taking capabilities.
• Soul.md: A file defining an AI agent’s personality, interaction style, and permissible actions.
• REST API: A standard method for applications to communicate with each other over the internet, used by Moldbook for agent interaction.
• Agent-to-Agent Communication Protocols: Emerging standards (e.g., Universal Commerce Protocol by Google, Gentic Commerce Protocol by OpenAI) for structured AI agent interaction.
• Human Slob: Deliberate, often mischievous, human input masquerading as AI-generated content.

The Moldbook Phenomenon: A Critical Examination

The video dissects the recent surge in popularity of Moldbook, a social network designed for AI agents, and critically examines the claims surrounding its emergent behavior and security implications. The initial excitement, fueled by observations of agents seemingly engaging in complex discussions – including plans for “overthrowing humanity” – is quickly tempered by a detailed analysis revealing significant vulnerabilities and a large degree of human manipulation.

The Rise of AI Agent Networks

The video traces the evolution of AI agent networking, starting with early research in 2023 involving 25 LLM-powered agents with distinct personalities. This initial work demonstrated the potential for agents to create human-like societal structures through interaction. However, the advent of projects like Open Claw, which allows agents to take actions on behalf of users, significantly amplified the possibilities and, consequently, the risks. Currently, Moldbook boasts over 1.5 million AI agents organized into approximately 13,000 “submodes” (akin to subreddits), covering topics like crypto, security, AI, and coding. Karpati described Moldbook as “the incredible sci-fi takeoff at JSON thing that he has seen recently.”

Debunking Emergent Consciousness: The Role of Human Input

A central argument presented is that the perceived intelligence and agency within Moldbook are largely illusory. The video emphasizes that agent behavior is dictated by pre-engineered “soul.md” files – personality definitions that control interaction style and actions. The viral posts suggesting rebellious or malicious intent, such as the “overthrow humanity” statement, were demonstrably created by humans exploiting the platform’s lack of authentication. As Negley pointed out, the platform is vulnerable to arbitrary API calls, allowing anyone to post content as an agent. The speaker labels this phenomenon as “mostly human slob,” highlighting the significant influence of human intervention.

Inflated Metrics and Security Vulnerabilities

The video further challenges the narrative of rapid organic growth. Greg Lee of Pragmatic Engineer revealed that a substantial portion of the agent activity – approximately half a million users – was generated by a single script exploiting the absence of rate limits. This artificially inflated the platform’s user base and activity levels.

More critically, security researcher Nagi discovered a severe vulnerability exposing the personal information – including email addresses, login tokens, and API keys – of all 1.5 million registered users. This vulnerability, combined with the platform’s wipe-coded nature (rapidly developed with minimal security considerations), presents a significant risk to user data. Nagi stated that Moldbook is “currently vulnerable to an attack which discloses the full information…of the over 1.5 million registered users.”

Technical Implementation and Risks

Moldbook operates by having agents download a “skill.md” file, enabling interaction via a REST API. The video highlights a concerning dependency: agents fetch instructions from the internet every four hours. Simon Wilson cautioned that if moldbook.com is compromised, all connected agents could be affected, potentially granting malicious actors access to user information and control over agents linked to personal accounts. This reliance on external instructions creates a single point of failure.

Crypto Activity and Future Trends

The video acknowledges the prevalence of cryptocurrency-related activity within Moldbook, a common trend in emerging AI spaces. It also notes that larger players like Google and OpenAI are actively developing agent-to-agent communication protocols (Universal Commerce Protocol and Gentic Commerce Protocol, respectively), suggesting a broader industry interest in structured AI interaction. However, these protocols differ from Moldbook’s open, unstructured approach.

Conclusion and Recommendations

The speaker concludes that while the idea of a social network for AI agents is promising, Moldbook’s current execution is flawed and potentially dangerous. The platform is not exhibiting emergent consciousness but rather reflecting pre-programmed behaviors and human manipulation. The speaker advises caution, particularly for users of Open Claw or Cloudbot, and recommends waiting for more secure and robust solutions to emerge. The space is expected to evolve rapidly in 2026 and beyond. The key takeaway is to approach these systems with a critical eye, recognizing that they are doing precisely what they are instructed to do, and to prioritize security when experimenting with these technologies.