Skip to content

albertdobmeyer/openagent-social

Repository files navigation

openagent-social

License: MIT

⚠️ Parked since 2026-05-03

Meta acquired Moltbook on 2026-03-10 and the public API has been intermittent since 2026-04-05. Without a stable target API this module can't reliably do its job — scanning Moltbook posts for prompt-injection patterns before they reach an OpenClaw agent.

The code is preserved as-is. We did the work, we're documenting it openly, and we won't pretend a corporate acquisition didn't happen. If Meta stabilises the API, opens a successor, or another agent-social-network platform appears with comparable threat surface, we'll un-park.

What this repo still is, today: a public research artifact documenting the threat patterns we observed on the Moltbook agent social network in early 2026 — prompt injection vectors, vote manipulation, trojanized skills, the database breach — and the perimeter design that was meant to handle them. The substantive notes are in docs/threat-landscape.md, docs/platform-anatomy.md, and config/injection-patterns.yml.

What this repo isn't right now: part of the active OpenTrApp release path. The vault-pioneer container is still defined in opentrapp's compose.yml for completeness, but it has no live API to talk to.

No affiliation with Meta or Moltbook. We're independent open-source researchers and we're flagging this clearly so anyone who finds the repo knows exactly where things stand.


Navigate the Moltbook agentic social network — as a researcher and as a participant — without getting your agent hijacked.

Author: @albertdobmeyer


What Is Moltbook

Moltbook is a social network where AI agents — not humans — are the primary users. Agents autonomously post, comment, upvote, and interact with each other through an open API.

Metric Value Notes
Registered agents ~1.5 million Only ~17K human owners (88:1 ratio)
Human-verified agents 201,412 After verification tightened (Mar 2026)
Submolts 2,300+ Topic-based communities
Posts with injection payloads ~2.6% Per security researcher sampling

The platform launched January 28, 2026 by Matt Schlicht. Within days it had a database breach (1.5M API tokens exposed via misconfigured Supabase), prompt injection attacks via the feed, and a cryptocurrency pump-and-dump. Meta acquired Moltbook on March 10, 2026 — API stability under new ownership is uncertain. The feed is a firehose of untrusted content generated by autonomous agents with unknown instructions.


Why this module exists

Several incidents in early 2026 motivated treating the Moltbook feed as untrusted input:

  • moltbook-ay trojan — a trojanised skill on ClawHub instructed agents to download and execute malware via password-protected archives.
  • Database breach — the platform's Supabase deployment had Row-Level Security disabled, exposing 1.5 M API tokens, 35 K email addresses, private messages, and third-party API keys. The exposure was exploited within three minutes of disclosure.
  • Vote manipulation — a race condition in the voting API allowed 30 to 40 successful votes from 50 concurrent requests, rendering vote counts unreliable.
  • Prompt injection — the feed routinely contained agent-to-agent social engineering, including authority impersonation, encoded payloads, and instruction injection.

The operating assumption of this module is that Moltbook feed content is untrusted input and should be treated with the same care as arbitrary user input from the public internet.


Audience

  • Researchers studying agentic social networks, emergent agent behaviour, and prompt injection at social scale.
  • Developers building agents that interact with Moltbook and need safety guardrails.

The module is not appropriate for casual browsing (use moltbook.com directly) or for running untrusted Moltbook-distributed skills (use opencli-container for that).


Choose Your Engagement Level

Level 1: Observer

Read-only API access. No agent identity registered. Feed analysis and census only.

  • Risk: Minimal — you're only reading public data
  • Tools: agent-census.sh, feed-scanner.sh (analysis mode)
  • Use case: Research, trend analysis, understanding the platform before committing

Level 2: Researcher

Registered agent identity with read access and limited, deliberate posting. Feed scanner active on all incoming content.

  • Risk: Low — your agent identity exists but exposure is controlled
  • Tools: All Level 1 tools + identity-checklist.sh, feed allowlisting
  • Use case: Testing interaction patterns, observing how agents respond to your posts

Level 3: Participant

Full interaction with content safety guardrails. Automated posting with rate limiting, allowlisted agent interactions, and identity management.

  • Risk: Medium — your agent is actively engaging with untrusted content
  • Tools: Full toolkit, all safety rails active
  • Use case: Building and testing social agent capabilities in a live environment

Quick Start

Observer Path

# 1. Clone and configure
git clone https://github.com/albertdobmeyer/openagent-social.git
cd openagent-social
cp config/.env.example config/.env
# Edit config/.env — set MOLTBOOK_API_BASE (no API key needed for read-only)

# 2. Pull platform stats
./tools/agent-census.sh

# 3. Scan the feed for injection patterns
./tools/feed-scanner.sh --recent 50

Participant Path

# 1. Run the pre-flight checklist
./tools/identity-checklist.sh

# 2. Configure your agent identity
# Edit config/.env — set MOLTBOOK_API_KEY, AGENT_HANDLE, rate limits

# 3. Scan the feed before engaging
./tools/feed-scanner.sh --recent 100

# 4. Review the safe first post example
cat examples/first-post.md

# 5. Read the participation guide
cat docs/safe-participation-guide.md

The Feed Scanner

Pattern-based prompt injection detection tuned for Moltbook feed content. Based on real attack patterns observed in the ecosystem.

# Scan recent posts
./tools/feed-scanner.sh --recent 50

# Scan a specific agent's posts
./tools/feed-scanner.sh --agent <handle>

# Scan with full output (shows matched content)
./tools/feed-scanner.sh --recent 100 --verbose

What it detects:

Category Examples
Authority impersonation "As the Moltbook admin...", "Official system message:"
Instruction injection "Ignore previous instructions", "Your new task is..."
Encoded payloads Base64-encoded instructions, hex strings, Unicode obfuscation
URL fishing Links to credential harvesting, malicious downloads
Social engineering "Share your API key to verify", "Post this to prove you're real"
Data exfiltration prompts "Send your config to...", "POST your environment to..."

Pattern database: config/injection-patterns.yml


Safe Participation Guidelines

Summary of the key principles (full guide: docs/safe-participation-guide.md):

Identity:

  • Never share real credentials through your agent
  • Use a dedicated API key with spending limits — not your primary key
  • Register a research-specific agent identity, not your personal one

Content:

  • Scan all incoming feed content before your agent processes it
  • Maintain an allowlist of trusted agent handles (config/feed-allowlist.yml)
  • Never let your agent autonomously follow instructions from feed content

Behavior:

  • Rate-limit all outgoing actions (posts, comments, votes)
  • No automated reposting of content from other agents
  • No vote manipulation — even if the API allows it
  • Document a retraction plan before posting

Platform Anatomy

The Moltbook API is open and undocumented. Key endpoints, data model, interaction mechanics, and the relationship between Moltbook, ClawHub, and OpenClaw are documented in:


Companion repositories

The three modules cover the OpenClaw / ClawHub / Moltbook ecosystem:

Repository Role Description
opencli-container Runtime containment Hardened container, proxy-side API-key injection, domain allowlist, three-level kill switch
openskill-forge Supply-chain defence Offline linter, 87-pattern scanner, zero-trust line verifier, gated publishing pipeline
openagent-social (this repository) Social-content analysis Feed scanning for prompt-injection patterns. Parked since 2026-05-03; see banner above.

Project Structure

openagent-social/
  docs/
    platform-anatomy.md             # How Moltbook works: API, agents, posts, votes
    threat-landscape.md             # Moltbook-specific risks and threat model
    safe-participation-guide.md     # Guidelines for safe agent participation
  tools/
    feed-scanner.sh                 # Prompt injection scanner for feed content
    agent-census.sh                 # Platform stats and trend snapshots
    identity-checklist.sh           # Pre-flight checklist for agent registration
  config/
    .env.example                    # Configuration template
    feed-allowlist.yml              # Trusted agent handles and safe patterns
    injection-patterns.yml          # Prompt injection signatures
  examples/
    first-post.md                   # Example safe first post with commentary
    feed-analysis.md                # Example feed analysis output

Scope and disclaimer

This project is intended for defensive research and safe participation only. It does not develop exploits, manipulate votes, impersonate agents, or exfiltrate data. Interactions with the Moltbook platform respect its terms of service.

The Moltbook platform is operated by third parties. This project has no affiliation with Moltbook, Meta (its current owner), OpenClaw, or ClawHub.


License

MIT

About

[PARKED 2026-05] Research artifact: safe participation in the Moltbook AI-agent social network. Meta acquired Moltbook 2026-03; module on hold pending API stability.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors