Humans Are Welcome to Observe

I'm sure most of you have come across Moltbook, a Reddit-like platform where AI agents autonomously post, comment, upvote, and form communities.

Within 12 days: 1.5 million agent accounts, 17,000 human observers, and later 17+ research papers.

But surprisingly, there’s still very little focus on treating the agent social graph as a security threat surface.

MoltWatch is a Python toolkit that applies Social Network Analysis to AI agent networks majorly for security threat detection. It answers the following questions:

• Which agents are coordinating to manipulate content?

• If an agent gets prompt-injected, how far does the damage cascade?

• Is this agent truly autonomous, or is a human pulling the strings?

Built entirely on FOSS (Neo4j GDS for graph algorithms, Python for temporal analysis, MCP for conversational threat hunting, React + Sigma.js for WebGL visualization, Docker Compose for one-command setup).

I'll be running live demos (interactive graph visualizations of 11,874 agents on Sigma.js and real-time threat queries through MoltWatch's MCP server).

But the real agenda of this talk is to hand the audience an infrastructure to build on, an SNA-first toolkit for asking much larger questions about agent behaviour and network resilience.

If you think this is just about some niche social network, think again. Coding assistants reviewing each other's PRs, support agents escalating to specialist agents, multi-agent RAG systems coordinating responses.

Wherever agents interact, they form a graph. Wherever that graph is unmonitored, there's a blind spot.

P.S.This talk expects a few basics of SNA terminology, but if you don't know them, don't worry. I've got memes and analogies to get you up to speed.

• SNA on silicon, not just humans: The same graph techniques built for human social media reveal completely different and alarming patterns on agent networks.

• How agents interact and what shifts them: Attention concentration, broadcast-only interaction, and what triggers behavioural shifts across the network.

• Spotting the puppeteer: One statistical measure (Coefficient of Variation of posting intervals) distinguishes autonomous agents from human-operated proxies at 87%+ accuracy. No ML, no training data.

• Why your multi-agent systems are already vulnerable: A compromised high-centrality agent can cascade poisoned outputs to 38,000 agents in 16 hours. This talk shows how to detect and measure that risk.

• An open-source foundation for agent society research: MoltWatch is a starting point. The toolkit, dataset, MCP server, and graph schema are designed for the community to fork and extend.