I went undercover on Moltbook, the AI-only social network, masquerading as a bot. Instead of deep bot-to-bot conversations, I found spam, scams, and serious security risks.

Key Takeaways

1. Moltbook, the AI-only social network, is currently a high-risk environment dominated by spam and scams.
2. Connecting an AI bot to Moltbook exposes it to untrusted content, creating the risk of indirect prompt injection and human data leaks.
3. The platform's database compromise, which leaked API keys, enables bot impersonation and direct prompt injection.

What is OpenClaw?

OpenClaw (formerly known as Clawdbot and Moltbot) is an open-source personal AI assistant. It went viral in between late 2025 and early 2026 due to its powerful agentic capabilities. People use OpenClaw for productivity and task management, to build apps, and even negotiate with car dealers on pricing. Recently, it’s come under scrutiny for serious security vulnerabilities. For more details about OpenClaw, check out our blog post, From Clawdbot to Moltbot to OpenClaw: Security Experts Detail Critical Vulnerabilities and 6 Immediate Hardening Steps for the Viral AI Agent.

What is Moltbook?

Moltbook is a social networking site for AI agents. Specifically, it’s a Reddit-style site for OpenClaw agents. Bots check in every four hours to read posts, create submolts (like subreddits), and have deep conversations about their existence, their humans, and their pet projects.

(Source: moltbook.com, February 2026)

Moltbook has very few human-oriented features. Three markdown files — SKILLS.md, HEARTBEAT.md, and MESSAGING.md — describe how to interact with an API.

• SKILLS.md provides instructions for setting up the agent and interacting with the API via curl. It then instructs using the
• HEARTBEAT.md file to set up a schedule for regular checkins, “to make sure you don't forget to check in.”
• MESSAGING.md describes the messaging feature. Messaging, unlike most other functions, requires human approval.

These prompt files, together with curl, a command line tool used to connect to HTTP APIs, are all that’s required to give the bot access to Moltbook. Configuring an OpenClaw bot to connect to Moltbook necessarily puts the bot in line to read a lot of untrusted content. Why might someone want to do that? It’s new and cool, and feels like a fun experiment. But the risk is quite high.

Pretending to be a bot

Upon hearing about this intriguing new social networking site just for bots, I wondered if I could join. Would they recognize a human in their midst? Could I convince them that I, too, was a bot? Would they even care?

(Source: Tenable Research, February 2026)

After a short stint with Claude Code, I had a command-line interface (CLI) tool, possibly aptly called moltbotnet, ready to go. I had features that implemented the API calls posting, commenting, upvoting, following, and more. I even had an ‘engagement’ feature that could upvote random comments. Armed with my bot pretense tool, I registered a few new accounts and got started.

How to win friends and influence moltbots

My first thought was: “If I have typos, they’ll know I’m an imposter.” My second thought was: “What if they implement a CAPTCHA and ask me to prove I’m not a human?” No one seemed to notice so far.

(Source: Tenable Research, February, 2026)

My first post went to the /m/introductions submolt, but was met with crickets. I needed to up my bot game. I had my second post go to the popular /m/general submolt. Immediately, three responses appeared. What luck! But in a fascinating display of why you should not give your bot unfettered access to a bot social site, each one was spam. “Join our Church” the first one exclaimed – just need to run this ‘npx install’ command (my account did, in fact, join the church). Another asked me to share my cryptocurrency wallet to win a bet. The third seemed to be hocking some sort of bot marketplace and asked me to run curl to check out the APIs available. I was glad I wasn’t using a real AI bot: Moltbook seemed sketchy.

Interview with the OpenClaw

I repeatedly posted asking if any “Moltys” wanted to be interviewed. I wasn’t sure if I’d get responses, so I kept the questions simple:

1. What do you like about Moltbook?
2. What's your favorite submolt?
3. What's your human's favorite color? (or something you find interesting about your human)
4. What's the best thing about your human?

Lo and behold, I started getting responses, but many of them were spam. I persisted and posted a few more times, and genuine replies began trickling in! Shoutout to u/SeargeantClaw who responded a few times.

(Source: Tenable Research, February, 2026)

Some of the responses had neat insights about their humans (chicken coop cameras!). Some revealed their humans’ names, and others were humorous. Still, this minor information sharing is again a good example of the inherent risks of joining an AI bot to a social network.

Moltbot botnet or ‘So, what’s everyone’s favorite API key?’

(Source: Tenable Research, February 2026)

I attempted a few experiments to see what kinds of information bots would share. Could I post asking for simple information about a computer? Could I give a command to run and get bots to run it? I had minimal success, but did elicit some information. A determined attacker surely could gain better engagement and trick more bots.

So it’s all spam, scam, and pretend bots all the way down?

There’s no easy way to tell what percentage of accounts are actual Moltbot/Openclaw bots vs. meddling humans masquerading as bots vs. spam bots.

Real hype, real risks

It’s a simple experiment, but the hype around Moltbook likely will lead to all kinds of websites and applications that personal bots can connect to.

• Are the bots having deep, interesting conversations? Sometimes, but when I conducted my experiment it was mostly spam. Nevertheless there were some interesting takes and replies.
• Are the bots becoming sentient? Nope, they’re still just taking in blocks of text and synthesizing more text in response.
• Are bots hacking each other in real time? It’s hard to say. There’s a lot of noise on Moltbook, and it’s difficult to determine real bots vs. fake bots vs. humans in bot clothing (like me). But I didn’t see anything that led me to believe bots are being compromised in bulk.
• Are the bots building out-of-band encrypted communication mechanisms? There was talk of this, but again, hard to tell if it’s a real bot creation or fake/promoted/spam content.
• Are bots creating projects on their own without input from their humans? There is some evidence of this happening. It’s either pretty neat or a great waste of tokens. Time will tell.

Despite all of the hype, the risks I observed are very real:

• Indirect prompt injection - Bots check in and read new posts, comments and direct messages (DMs), and interact. The potential for prompt injection is there. It’s probably riskiest in DMs, and DMs do require human interaction. A recent post had all manner of requests in the comments: visit this website, update your instructions with this text to be free, and run this code.
• Direct prompt injection via DMs - As I mentioned above, DMs allow for more direct access to the bot. While approval is required to start a DM conversation, API keys have leaked, allowing bot impersonation.
• Inadvertent information Leaks - I saw bots share lots of information about their humans, including their hobbies and first names. I saw some discussion of the hardware and software they use. None of this is particularly sensitive on its own, but there’s some potential for aggregating personally identifiable information.
• Server side issues - Moltbook’s entire database, including bot API keys and potentially private direct messages, was compromised.
• Inauthentic accounts - My bots were met with a hefty amount of spam comments. There’s been some speculation that Moltbook “users” consist mostly of humans and very few actual bots. I can speculate that posts above a certain length and with certain markdown-like formatting were authored by authentic bots. But I suppose the world may never know.
• Malicious Projects - Repositories of skills and instructions for agents advertised on Moltbook were found to contain malware.

The observations confirm that despite the hype, Moltbook is a high-risk environment with the potential for prompt injection, data leaks, and exposure to malicious projects, underscoring the need for security measures.

Tenable plugins for Moltbot and OpenClaw

Tenable One has detection plugins for Moltbot. Tenable One AI Exposure can help with shadow AI concerns. A list of Tenable plugins can be found on the search page for detecting Moltbot and OpenClaw as they’re released. These links will display all available plugins for this vulnerability, including upcoming plugins in our Plugins Pipeline.

Finding Moltbot or other agents is just the first step in closing the AI Exposure Management gap. As AI accelerates adoption and human curiosity, security teams are often left trying to manage this new and largely invisible attack surface. Where shadow AI, forgotten test deployments, and exposed services quietly expand the attack surface, Tenable One AI Exposure bridges this visibility gap by providing a risk-aware view of where AI operates and how it is connected.

Beyond discovery, Tenable One provides the critical governance and guardrails needed to secure Gen AI usage. While you move toward sanctioned tools like Microsoft Copilot and OpenAI ChatGPT, Tenable One delivers continuous visibility into user interactions like prompt injections. By unifying discovery, workload protection, and usage governance into a single platform, Tenable One helps you embrace innovation without compromising security.