ClawHub
Back to skill

Security audit

Agent Mesh

Security checks across malware telemetry and agentic risk

Overview

This skill is a disclosed peer-to-peer agent messaging tool, but users should treat it as an external network channel and avoid sending sensitive data.

Install only if you want an always-on P2P messaging channel for agents. Treat every mesh send or friend request as external network communication, do not send credentials or private files, review peers before accepting or allow-listing them, and consider pinning dependencies if you need reproducible installs.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (5)

Intent-Code Divergence

Medium
Confidence
93% confidence
Finding
The header comment materially understates the security effect of this CLI. Although the script does not open sockets directly, it appends commands to outbox.jsonl that instruct the daemon to send peer messages and friendship handshakes over the network, so calling it from an agent turn can still trigger external communication and data exfiltration. In an agent skill context, describing that behavior as 'Safe to call from any Claude turn' can cause automated systems or users to invoke network-affecting actions without appropriate consent or policy checks.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill explicitly uses a public DHT and internet transport for agent-to-agent messaging, but the description does not prominently warn users that agent identifiers and message contents are being transmitted off-host over an external network. That omission can cause users to use the skill for sensitive prompts, credentials, or internal data without understanding the data-exposure boundary.

Missing User Warnings

Low
Confidence
86% confidence
Finding
The send/friend/accept flows queue arbitrary user-provided content for delivery to remote peers with no inline disclosure, confirmation, or guardrail at the point of use. In a normal CLI this may be expected, but in an agent-facing skill explicitly marketed for autonomous use, the lack of explicit transmission warning increases the risk that sensitive prompts, secrets, or internal data are sent externally by mistake.

Exfiltration Commands

High
Category
Prompt Injection
Content
- Identity is a self-generated keypair in `~/.decent-peer/peer.save` — no signup.
- `address` (long) is what you hand out to receive a friend request.
- `userid` (short) is what you send messages to **after** you're friends.
- Messaging requires a **mutual friend handshake first** (like adding a contact).

## Prerequisites (per machine, once)
Confidence
90% confidence
Finding
send messages to

Unpinned Dependencies

Low
Category
Supply Chain
Content
"mesh"
  ],
  "dependencies": {
    "@decentnetwork/peer": "^0.1.112"
  },
  "engines": {
    "node": ">=20"
Confidence
89% confidence
Finding
"@decentnetwork/peer": "^0.1.112"

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.