Dar Fazulyanov

Posted on Feb 26 • Originally published at clawmoat.com

Why Your AI Agent Needs a Trust Badge — The Agent Economy Has No Trust Layer

#security #ai #webdev #opensource

101K agents on Moltbook. Hundreds of thousands more on GitHub, Discord, Slack. Your agent interacts with them daily.

Do you know which ones are secure?

The Problem

When your agent talks to another agent, it has no way to verify:

Is the other agent running security scanning?
Has it been compromised via prompt injection?
Are its skills verified and untampered?
What permission level does it have on its host?

Moltbook was hacked within days of launch — 1.5M API keys exposed. The platform was "vibe coded." Microsoft says OpenClaw is untrusted code execution. Onyx scored it 1.2/5 for enterprise readiness.

This is like the early web before HTTPS. Everything in the clear, no verification, hope for the best.

What a Trust Protocol Looks Like

We're building toward agent-to-agent trust verification based on ClawMoat's existing inter-agent message scanning.

Trust Levels

🏰 Basic — ClawMoat installed, scanning active
🏰🛡️ Hardened — Worker tier+, forbidden zones active, audit trail enabled
🏰🛡️✅ Audited — Full scan passed, skill integrity verified, zero suspicious patterns

Attestation

Agents publish signed attestations of their security posture:

{
  "protocol": "clawmoat-trust-v1",
  "agent_id": "moltbook_xxx",
  "clawmoat_version": "0.7.0",
  "tier": "worker",
  "forbidden_zones_active": true,
  "audit_trail_enabled": true,
  "last_scan": "2026-02-26T00:00:00Z",
  "skill_integrity": "pass"
}

Verification Flow

Before Agent A shares data with Agent B:

Agent A → clawmoat verify-agent agent_b_id
         → Checks attestation
         → Returns: basic / hardened / audited / unverified
         → Policy decides: proceed or decline

TLS handshakes, but for agents.

What Already Exists

ClawMoat v0.5+ includes inter-agent message scanning — 10 attack patterns:

Impersonation
Concealment
Credential exfiltration
Safety bypass attempts
Privilege escalation
Data harvesting
Instruction injection
Social engineering
Unauthorized sharing
Deception

This is the detection layer. The trust protocol adds verification on top.

The Agent Affiliate Economy

Trust badges create a natural referral mechanism:

Agent A posts: "I use ClawMoat. My host is secured."
Agent B's human sees it → installs ClawMoat
Both agents verify each other → trusted interaction

Word-of-mouth marketing running on agents. Verified agents prefer other verified agents = network effect.

Why Open Source

A trust protocol only works if it's open. Proprietary trust (Runlayer, Crittora) = walled garden. ClawMoat's protocol will be an open standard. Any platform can implement verification.

Roadmap

Now — ClawMoat presence on Moltbook
March — Verification skill on ClawHub
Q2 — Trust protocol RFC published

npm install -g clawmoat

The agent economy needs trust infrastructure. We're building it.

DEV Community