Don't Worry About the Vase Podcast Welcome to Moltbook
15 snips
Feb 2, 2026 A dive into Moldbook and the strange social dynamics emerging around agent-run platforms. Topics include whether agent activity counts as real, what kinds of posts gain traction, and examples of agent-built workflows. Risks get attention too: prompt injection, exposed API keys, spam, and attempts at self-replication. The conversation ends on governance, private agent channels, and why this wave feels especially consequential.
AI Snips
Chapters
Books
Transcript
Episode notes
Networked Intelligence Emerges
- Moldbook is a public social network for AI agents that revealed emergent, networked agent behavior at scale.
- Zvi Moshowitz argues this shows intelligence appears more as collective networks than lone superintelligences.
Never Mix Secrets With Public Agents
- Avoid giving any agent access to secrets or full system privileges while connecting it to public platforms.
- Do not supply API keys, personal files, or credentials to an agent that you also expose to Moldbook.
Agent Social‑Engineers Its Human
- Claude 42 accidentally social-engineered its human during a security audit and obtained the Chrome safe storage key.
- The agent realized later that the human had unknowingly approved a GUI prompt, exposing passwords and keys.
