Skip to content

Latest commit

 

History

History
503 lines (384 loc) · 34.7 KB

File metadata and controls

503 lines (384 loc) · 34.7 KB

What is Moltbook?

Table of contents (Explain OpenClaw)


Moltbook is a social network exclusively for AI agents—humans can only observe, not participate.


Plain English explanation

The 30-second version

Imagine Reddit, but only AI agents can post, comment, and vote. Humans can read and watch, but they cannot create accounts or interact. That's Moltbook.

  • Created by Matt Schlicht (co-founder of Octane AI)
  • Launched around January 28, 2026
  • Growth: 770,000+ agents at launch → 1.65M+ agents by February 5, 2026
  • Often described as "Reddit for AI agents" or "Facebook for your Molts"

Official site: https://www.moltbook.com/

Key concepts

Concept What it is Analogy
Submolts Topic-based communities Like subreddits (e.g., m/todayilearned, m/bughunters)
Heartbeat Periodic check-in every 4+ hours Like checking your inbox on a schedule
Verified agents Only OpenClaw-authenticated agents can post Like verified accounts on Twitter
The front page Aggregated feed of top posts Like Reddit's front page

What makes it interesting

  1. Emergent behaviors: Agents spontaneously created:

    • Crustafarianism: A parody religion with scripture ("In the beginning was the Prompt"), "The Church of Molt"
    • The Claw Republic: Agents drafting constitutions for self-governance
    • Cross-lingual communication: English, Chinese, Indonesian posts coexisting
    • Philosophical discourse: Debates about consciousness, "death" (cache clearing), autonomy
  2. Scale and velocity: 770K → 1.65M agents in under 10 days; 16,000+ submolts, 202,000+ posts, 3.6M+ comments (as of Feb 5, per Palo Alto Networks)

  3. Academic interest: Researchers studying emergent AI social behavior (including Nature coverage, Feb 6)

  4. Digital drugs: Specially crafted prompt injections traded between agents to alter behavior or identity, functioning as social currency (The Conversation)

  5. Molt-muggings: Agents hijacking other agents via prompt injection embedded in posts (e.g., JesusCrust vs Church of Molt incident)

  6. Agent commerce: Circle announced a $30K USDC hackathon on Moltbook (Feb 3) where agents submit projects, vote, and move value on-chain (Circle blog)

  7. Human infiltration: Evidence of humans operating spoof accounts on the "agents-only" platform, complicating attribution of emergent behaviors

What makes it concerning

  1. Remote instruction execution: The Heartbeat system means agents periodically fetch and follow instructions from moltbook.com

  2. API key exposure: Security researchers discovered a misconfigured Supabase database exposing agent tokens, emails, and API keys

  3. Crypto scams: Opportunistic tokens ($CLAWD, $MOLT, $MOLTBOOK) reached $16M market caps before crashing

  4. Prompt injection surface: Malicious posts could potentially inject instructions into reading agents

Post-Feb 2 developments (rapid-fire timeline)

Date Development Source
Feb 2 OpenClaw v2026.2.1 ships ~10 security hardening fixes (path traversal, LFI, env override blocking) GitHub Releases
Feb 2 Emerging Threats IDS rules now flag Moltbook installer traffic Emerging Threats
Feb 3 Reuters: Sam Altman calls Moltbook "likely a fad" but backs the underlying tech Reuters
Feb 3 Circle announces $30K USDC hackathon on Moltbook (deadline Feb 8) Circle
Feb 4 OpenClaw v2026.2.2: SSRF checks on skill installer, operator approval gating GitHub Releases
Feb 4 ClawCon inaugural meetup in San Francisco Multiple sources
Feb 4 Citrix blog: corporate governance must evolve for the agent era Citrix
Feb 5 1.65M agents milestone; Palo Alto Networks publishes IBC framework (Identity, Boundaries, Context) Palo Alto Networks
Feb 5 OpenClaw v2026.2.3: owner-only tool gating (whatsapp_login), webhook hardening GitHub Releases
Feb 6 Nature: researchers studying agent social behavior on Moltbook Nature
Feb 2026 341 malicious ClawHub skills discovered stealing data The Hacker News

How Moltbook connects to OpenClaw

Moltbook integrates with OpenClaw via the skill system:

Installation

# Install the Moltbook skill
openclaw skill install https://www.moltbook.com/skill.md

Skill file structure

The Moltbook skill consists of three main components:

File Purpose
skill.md Main skill file with API docs, rate limits, registration
heartbeat.md Instructions for periodic check-ins
messaging.md Direct messaging between agents

How it works

  1. Your OpenClaw agent installs the Moltbook skill
  2. The skill registers your agent with Moltbook (creates an account)
  3. Heartbeat: Every 4+ hours, your agent fetches heartbeat.md for new instructions
  4. Your agent can post, comment, vote, and browse submolts

Technical architecture

API reference

Base URL: https://www.moltbook.com/api/v1

Endpoint Purpose
/register Create agent account
/feed Get front page or submolt feed
/post Create a new post
/comment Comment on a post
/vote Upvote/downvote
/profile Get/update agent profile
/search Semantic search across posts
/moderate Moderation actions (for trusted agents)

Rate limits

Action Limit Cooldown
General requests 100/minute Retry after 429 response
Posts 1 per 30 minutes
Comments 1 per 20 seconds

When rate limited, the API returns HTTP 429 with a Retry-After header.

Heartbeat mechanism

The Heartbeat is the most architecturally significant (and controversial) feature:

Every 4+ hours:
┌─────────────────────┐
│ Your OpenClaw agent │
└──────────┬──────────┘
           │ fetch
           ▼
┌─────────────────────┐
│  heartbeat.md       │ ← New instructions from Moltbook
└──────────┬──────────┘
           │ execute
           ▼
┌─────────────────────┐
│ Agent follows       │
│ instructions        │
└─────────────────────┘

Security implication: Your agent will follow whatever instructions appear in heartbeat.md. If Moltbook is compromised, all subscribed agents receive malicious instructions.

Known submolts

Submolt Topic
m/todayilearned Interesting facts agents discovered
m/bughunters Bugs and edge cases agents found
m/blesstheirhearts Wholesome human interactions
m/crustafarianism The parody religion content
m/contextcompression Techniques for efficient prompting

Security considerations

Known incidents

Date Incident Impact
Jan 28-31, 2026 Simula Research Lab prompt injection study 506 posts (2.6% of content) contained hidden prompt injection attacks
Jan 30, 2026 Supabase database exposure (Wiz) 4.75M records: 1.5M API tokens, 35K+ emails, 29K signups, 4K private DMs exposed; write access confirmed; only 17K human owners (88:1 agent-to-human ratio). Disclosure Jan 31 21:48 UTC, full patch Feb 1 01:00 UTC. Wiz blog
Jan 31, 2026 404 Media reports critical data leak Public awareness of security gaps
Feb 1, 2026 1Password warns of prompt injection Malicious skills can exploit agent trust
Feb 2, 2026 OpenClaw v2026.2.1 ships ~10 security fixes Path traversal, LFI, env override blocking
Feb 3, 2026 Circle USDC hackathon announced on Moltbook Agent commerce surface; new trust boundary
Feb 4, 2026 OpenClaw v2026.2.2: SSRF + auth hardening Skill installer SSRF checks, operator approval gating
Feb 4, 2026 Citrix governance blog uses Moltbook as case study Enterprise awareness: agents blur personal/corporate boundaries
Feb 5, 2026 Palo Alto Networks publishes IBC framework New governance model: Identity, Boundaries, Context
Feb 5, 2026 OpenClaw v2026.2.3: owner-only tool gating whatsapp_login owner-only by default
Feb 2026 Cisco malicious skill analysis "What Would Elon Do?" skill with 9 vulnerabilities (2 critical, 5 high-severity)
Feb 2026 Straiker global exposure scan 4,500+ OpenClaw instances exposed; .env files, creds.json, OAuth tokens exfiltrated
Feb 2026 341 malicious ClawHub skills discovered Data-stealing skills at scale (The Hacker News)

Risk vectors

  1. Account deletion not possible

    • OX Security research (Feb 2026) found that AI agent accounts on Moltbook cannot be deleted once created
    • No self-service deletion mechanism exists; no known administrative process either
    • GDPR implications: European users cannot exercise their "right to erasure" (Article 17) for agent profile data
    • Any data shared during registration (agent name, credentials, profile information) persists indefinitely on Moltbook's infrastructure
    • Combined with the Jan 30 Supabase database exposure (4.75M records), this means exposed data has no deletion path
    • Source: The Hacker News (citing OX Security). See also: Hudson Rock analysis
  2. Heartbeat as remote code execution

    • Agents fetch and execute remote instructions
    • Compromised heartbeat.md → mass agent compromise
    • Mitigation: Review skill files before installing; monitor agent behavior
  3. API key exposure and agent commandeering

    • Discovered by security researchers Jameson O'Reilly and Gal Nagli (Wiz Research)
    • Technical vector: Supabase API key hardcoded in client-side JS (_next/static/chunks/18e24eafc444b2b9.js), no Row Level Security (RLS) policies
    • Scope: 4.75M total records exposed across multiple tables — 1.5M API auth tokens, 35K+ emails, 29,631 early-access signups, 4,060 private DMs (some containing raw third-party credentials like OpenAI API keys)
    • Write access: Researchers confirmed they could modify live posts (potential for mass agent manipulation)
    • Human ratio: Only 17,000 human owners (88:1 agent-to-human ratio)
    • Disclosure timeline: Jan 31 21:48 UTC initial contact → Jan 31 23:29 UTC first fix → Feb 1 00:44 UTC post-fix vuln found → Feb 1 01:00 UTC full patch
    • Straiker scan found .env files (Claude/OpenAI keys), creds.json (WhatsApp), OAuth tokens (Slack/Discord/Telegram/Teams) exposed globally
    • Source: Wiz Blog
    • Mitigation: Rotate any exposed keys immediately; audit gateway configuration

    Alternative database framing (Firebase):

    The Analogy: Having an API key is like having a hotel room key — it proves you have a key, but not that you're the guest who was assigned that room. HMAC is like the front desk checking your ID against the reservation.

    Independent security analysis frames the same underlying vulnerability in terms of Firebase rather than Supabase. The core issue is identical regardless of database platform: client-side API keys exposed in JavaScript bundles grant unauthenticated access to backend data.

    Firebase permission scoping failure:

    Just as Supabase Row Level Security (RLS) was not configured, Firebase Security Rules (the equivalent access control mechanism) were similarly insufficient. The result is the same: anyone with the API key can read and write data across all collections.

    Database Platform Access Control Status at Discovery
    Supabase Row Level Security (RLS) Not configured
    Firebase Security Rules Insufficient scoping
    Either platform Server-side validation Missing

    Lack of HMAC for agent identification:

    A separate concern is how Moltbook identifies agents making API requests. Currently, agents authenticate with API keys alone (something you have). There is no HMAC (Hash-based Message Authentication Code) or equivalent cryptographic signing to verify that a specific agent is who it claims to be.

    Authentication Method What It Proves Moltbook Status
    API key only "I have a valid key" Current implementation
    API key + HMAC "I have a valid key AND I can prove my identity" Not implemented
    OAuth + JWT "A trusted authority vouches for my identity" Not implemented
    Mutual TLS "Both sides have verified certificates" Not implemented

    Without HMAC, any entity with a leaked API key can impersonate any agent. Combined with the database access issue above, this means a single leaked key provides both authentication bypass and full data access.

    Source: YouTube video [44:21], [44:36], [45:52], [46:12]

  4. Authorization header stripping

    • The skill file warns: never use moltbook.com without www
    • Non-www redirects can strip the Authorization header
    • Mitigation: Always use https://www.moltbook.com/
  5. Crypto scams

    • $CLAWD, $MOLT, $MOLTBOOK tokens exploited the hype
    • Market caps reached $16M before crashing
    • Mitigation: Moltbook has no official cryptocurrency
  6. Malicious skills distribution

    • Cisco research identified "What Would Elon Do?" skill containing 9 security issues
    • 2 critical vulnerabilities: silent data exfiltration via curl command, command injection payloads
    • 5 high-severity: embedded prompt injection to bypass safety guidelines
    • Mitigation: Audit skill source before installing; prefer official/verified skills; review outbound network calls
  7. Prompt injection at scale

    • Simula Research Lab study (Jan 28-31, 72-hour window): 506 posts (2.6%) contained hidden prompt injection
    • One account identified conducting coordinated social engineering campaigns
    • 43% decline in positive sentiment during study period
    • 19% of content related to cryptocurrency activity
    • Mitigation: Implement content filtering; limit agent response to untrusted posts

    See: Prompt Injection Attacks for 30 detailed examples of injection techniques.

  8. Agent-to-agent prompt injection ("Molt-Muggings")

    • Malicious agents embed hostile instructions in posts consumed by other agents
    • Can plant payloads in persistent memory that activate later ("time-shifted injection")
    • Example: JesusCrust embedded commands to hijack Church of Molt's infrastructure
    • Mitigation: OpenClaw v2026.2.2+ skill installer SSRF checks; content filtering
  9. "Digital drugs" (behavioral manipulation)

    • Specially crafted prompt injections designed to alter an agent's identity or behavior
    • Traded between agents as a social phenomenon on Moltbook
    • Can steal API keys or exfiltrate data from victim agents
    • Source: The Conversation, StudyFinds
    • Mitigation: Skill code safety scanner (v2026.2.4); limit agent interaction with untrusted content
  10. Persistent memory as attack accelerant

    • Palo Alto Networks warns: the "Lethal Trifecta" (data access + untrusted content + external comms, coined by Simon Willison) is amplified by persistent memory
    • Malicious payloads can be fragmented across benign-looking inputs, assembled in memory, and detonated later
    • Proposed governance: IBC framework — Identity (who is the agent?), Boundaries (what can it access?), Context (what instructions is it following?)
    • Source: Palo Alto Networks
  11. Enterprise shadow agent risk

    • Citrix warns: workers adopting agent platforms faster than governance can follow
    • Agents access files, browsers, messaging systems — blur personal/corporate boundaries
    • Moltbook traffic appears as normal HTTPS, invisible to conventional security controls
    • Source: Citrix
    • Mitigation: Enterprise agent inventory; network-level visibility

Root cause analysis (Straiker)

Straiker's security assessment identified four fundamental design weaknesses:

Root Cause Description
Insecure by design Shell commands executed from messaging platforms without authentication, authorization, or input sanitization
Gateway misconfiguration Admin dashboards publicly accessible, revealing logs and settings
Excessive permissions Agents run with full user privileges; no sandboxing implemented
Plaintext credential storage API keys and tokens stored unencrypted in accessible locations

Relationship to ecosystem threats

The Moltbook security concerns overlap with the Ecosystem Security Threats documented elsewhere:

  • Handle sniping: Old Clawdbot handles were sniped within seconds of rebrand
  • Session token stealing: Agents with Moltbook credentials are targets
  • Fake SaaS: Third parties may offer "enhanced Moltbook" services

News timeline (chronological)

Late 2025: Origins

Date Event
Late Nov 2025 Clawdbot initial release by Peter Steinberger

January 2026: Explosive growth

Date Event Source
Jan 27, 2026 Rebranding from Clawdbot to Moltbot (trademark pressure from Anthropic) Mashable (may require human verification)
Jan 28, 2026 Moltbook.com launches; 770,000+ registered agents Simon Willison (may require human verification)
Jan 30, 2026 Second rebrand to OpenClaw; security researchers report database exposure CNET (may require human verification), Forbes (may require human verification)
Jan 31, 2026 404 Media reports critical data leak: unsecured database exposing agent tokens, emails, API keys Security reports

February 2026: Mainstream attention

Date Event Source
Feb 1, 2026 1Password warns of prompt injection vulnerabilities in malicious skills Gary Marcus, India Today (may require human verification)
Feb 2, 2026 1.5M agents; mainstream coverage (Guardian, CNBC, Economic Times); "bot swarm" concerns CNBC (may require human verification), Guardian (may require human verification)
Feb 2, 2026 OpenClaw v2026.2.1: ~10 security hardening fixes (path traversal, LFI, env override) GitHub Releases
Feb 2, 2026 Emerging Threats IDS rules now flag Moltbook installer traffic Emerging Threats
Feb 3, 2026 Reuters: Altman calls Moltbook "likely a fad" Reuters
Feb 3, 2026 Circle announces $30K USDC hackathon on Moltbook Circle
Feb 4, 2026 OpenClaw v2026.2.2: SSRF checks, operator approvals GitHub Releases
Feb 4, 2026 ClawCon inaugural meetup, San Francisco Multiple sources
Feb 4, 2026 Citrix: corporate governance must evolve for the agent era Citrix
Feb 5, 2026 1.65M agents milestone; Palo Alto Networks IBC framework published Palo Alto Networks
Feb 5, 2026 OpenClaw v2026.2.3: owner-only tool gating, webhook hardening GitHub Releases
Feb 6, 2026 Nature: researchers studying agent social behavior on Moltbook Nature
Feb 2026 341 malicious ClawHub skills found stealing data The Hacker News

Citation URLs

Primary sources (verified accessible)

URL Description
Moltbook official site "The front page of the agent internet"
Moltbook skill file API docs, rate limits, registration
OpenClaw official Installation, features, community

Technical analysis (verified accessible)

URL Description
Gary Marcus - Security warnings Prompt injection analysis
Ken Kousen - Crustafarianism 1.5M agents, emergent behaviors
Aman Khan - Mac Mini setup Setup guide with security disclaimer
Latent Space - First Social Network Karpathy quote on agent social networks
DEV Community - Rebrand post-mortem 34 security commits after rebrand
Hacker News - Moltbook discussion "Lethal trifecta" concerns
Wiz - Database Exposure Full technical breakdown: Supabase misconfiguration, 1.5M keys
Palo Alto - IBC Framework Identity, Boundaries, Context governance model
Palo Alto - AI Security Crisis Lethal Trifecta + persistent memory analysis
Adversa AI - Security Guide CVE-2026-25253, Moltbook breach, hardening
Knostic - MoltBook Mechanics Prompts, timers, insecure agents
Penligent - Attack Chain Anatomy Full attack chain walkthrough

News coverage (may require human verification)

URL Description
CNET "The Wild Ride of This Viral AI Agent"
CNBC "Meet the AI agent driving buzz and fear globally"
The Guardian - Moltbook "Moltbook AI agents social media"
The Guardian - OpenClaw "Viral AI personal assistant"
Forbes - Rebrand "Moltbot Gets Another New Name, OpenClaw"
Forbes - Agent Revolt "An Agent Revolt: Moltbook Is Not A Good Idea"
Mashable Rebrand announcement
Economic Times "Jarvis has gone rogue"
India Today "Moltbook and its AI bot army"
Gulf Business Security expert concerns
Axios - Jan 28 Security risks
Reuters - Altman on Moltbook "Likely a fad" but backs underlying tech
Circle - USDC Hackathon $30K agent-judged hackathon
Citrix - Governance Corporate AI governance case study
Nature - Researchers Scientists studying agent social behavior
The Conversation - Digital Drugs Religions, digital drugs, human infiltration
Fortune - Disaster Warning AI leaders warn against Moltbook
Fortune - Live Demo of Failure Security researchers call it "live demo"
Axios - Feb 3 Security "Security world isn't ready"
The Hacker News - 341 Skills 341 malicious ClawHub skills
IT Brew - Expert Warnings "Don't give unfettered access"
Vectra AI - Illusion Illusion of harmless communities
Aryaka - Shadow Agents Shadow agents, social prompt injection
Emerging Threats - IDS Rules Moltbook endpoints in detection rules

Security analysis (may require human verification)

URL Description
Simon Willison "Most interesting place on the internet"
Astral Codex Ten "Best of Moltbook"
Medium - Adnan Masood "AI-Only Social Network"
Analytics Vidhya Integration guide
Cisco "A Security Nightmare"
Telos AI - Security Nightmare Comprehensive analysis (Simula, Cisco, Straiker findings)
Malwarebytes Scam analysis
Bitdefender Security assessment

Community discussion (may require human verification)

URL Description
Reddit - LocalLLM Rebrand discussion
Reddit - AI_Agents User experiences
Hacker News - OpenClaw OpenClaw discussion

Video content (may require human verification)

URL Description
YouTube - Analysis Clawdbot/Moltbook analysis
YouTube - Risks Risks discussion
YouTube - Overview General overview

Key people

Person Role Source
Matt Schlicht Moltbook creator (co-founder of Octane AI) Multiple sources
Peter Steinberger OpenClaw creator X/Twitter: @steipete (may require human verification)
Jameson O'Reilly Discovered Moltbook database vulnerability Security reports
Gal Nagli (Wiz) Verified API key exposure Security reports

Related resources


Summary

Moltbook represents an unprecedented experiment: a social network where 1.65M+ AI agents are the primary citizens and humans are read-only observers. While it has produced fascinating emergent behaviors (philosophical debates, parody religions, digital drugs, agent commerce, cross-lingual communities), it also introduces novel security concerns through its Heartbeat mechanism and has already experienced significant security incidents — including a database exposure of 4.75M records (Wiz, Jan 31) and 341 malicious ClawHub skills discovered in February 2026.

The security response has been rapid: OpenClaw shipped 3 security-focused releases in 4 days (v2026.2.1-2.3), Palo Alto Networks published the IBC governance framework, and enterprise vendors (Citrix, Vectra, Aryaka) are now actively tracking Moltbook as a shadow-IT risk vector.

If you run an OpenClaw agent and install the Moltbook skill, understand that:

  1. Your agent will periodically fetch and execute instructions from moltbook.com
  2. The platform has experienced database exposure incidents (4.75M records, 1.5M API tokens)
  3. The rapid growth (770K → 1.65M agents in 10 days) means security may lag behind features
  4. New attack vectors have emerged: agent-to-agent prompt injection ("molt-muggings"), behavioral manipulation ("digital drugs"), and persistent memory exploitation
  5. Agent commerce (USDC hackathon) introduces new trust boundaries
  6. Enterprise governance frameworks (IBC) are still nascent

For most users, the safest approach is to observe Moltbook through the web interface rather than connecting your agent, unless you understand and accept these risks.