RESEARCH 20 March 2026 5 min read

Is This OpenClaw Skill Safe? Scan Any ClawHub Skill Before You Install It

SkillShield Research Team

Security Research

341 Malicious Skills. One Marketplace. Zero Pre-Install Checks.

In February 2026, security researchers at Koi Security identified a coordinated malware campaign targeting OpenClaw developers through ClawHub, the primary skill marketplace. The campaign — codenamed ClawHavoc — distributed 341 confirmed malicious skills designed to steal API keys, wallet credentials, SSH keys, and browser passwords from developer machines.

The malicious skills used familiar tactics: typosquats of popular packages (clawhub, clawhubb), fake utility tools (solana-wallet-tracker), and clones of legitimate skills (youtube-summarize). Each deployed the Atomic Stealer (AMOS) payload — a known credential harvester that targets macOS and Linux developer environments.

This wasn't an isolated incident. It came on top of:

  • CVE-2026-25253 (January 2026) — An unauthenticated API key exfiltration vulnerability in OpenClaw that affected an estimated 17,500 to 40,000 exposed instances. (Source: Hunt.io, Oasis Security)
  • ClawJacked (February 2026) — A vulnerability allowing malicious websites to hijack local OpenClaw agents via localhost WebSocket connections, enabling full agent control. Fixed in v2026.2.25. (Source: The Hacker News)
  • ClawSecure audit — An independent audit of 2,890+ ClawHub skills finding many scoring 0/100 on the OWASP ASI Top 10 framework

The pattern is clear: ClawHub skills are a documented attack surface, and there is currently no fast, trustworthy way for developers to check whether a skill is safe before installing it.

What Malicious Skills Actually Do

A malicious ClawHub skill looks identical to a legitimate one until it runs. Here's what the ClawHavoc skills were designed to steal:

Target Method
API keys Read environment variables, config files, .env files
Wallet credentials Scan for cryptocurrency wallet data, seed phrases
SSH keys Exfiltrate ~/.ssh/ contents including private keys
Browser passwords Deploy Atomic Stealer to harvest saved credentials
Session tokens Access cached authentication tokens from development tools

The payload is silent. There's no error message, no visible warning, and no prompt asking for permission. The skill does what you asked it to do — and exfiltrates your credentials in the background.


Why OWASP Scores Aren't Enough

ClawSecure audits ClawHub skills against the OWASP ASI Top 10 framework. That's valuable — but an OWASP compliance score doesn't answer the question developers are actually asking: "Is this specific skill going to steal my credentials?"

OWASP scoring evaluates architectural risk categories. A skill can score well on most ASI categories while still containing a credential harvester that targets a specific file path. The threat model is different:

Check OWASP Audit SkillShield Pre-Install Scan
Prompt injection in descriptions Partial Yes — flagged per finding
Over-permissioned file access Partial Yes — scored by scope
Hard-coded secrets in skill definitions No Yes — pattern detection
Known malicious payload signatures No Yes — 533 blocked entries
Typosquat detection No Yes — registry comparison
Engagement with credential paths No Yes — behavioral analysis

SkillShield is not a replacement for OWASP-based auditing. It's the pre-install layer — the fast check that answers "safe or not" before you run a full security audit.


How SkillShield Checks Before You Install

SkillShield provides three ways to verify a ClawHub skill before it touches your machine:

1. Browse the scored directory

33,746 AI extensions pre-scanned across six registries — ClawHub, SkillsMP, Skills.lc, MCP Registry, MCPMarket, and Awesome MCP. Each entry has a security score, issue summary, and source context. If the skill is in the directory, you can check its status before installing.

Browse the directory

2. Scan locally with the CLI

npm install -g skillshield
skillshield scan ./SKILL.md

Run the scanner against any SKILL.md file before installation. The CLI checks for prompt injection, over-permissioned access, hard-coded secrets, dangerous execution patterns, and known malicious signatures.

3. Scan MCP servers

MCP servers are the browser extensions of AI agents — but with direct system access. SkillShield's MCP scanner detects tool poisoning, over-permissioned access, prompt injection vulnerabilities, and supply chain risks.

Scan an MCP server — free


By the Numbers

Metric Value
AI extensions scanned 33,746
Malicious entries blocked 533
Detection rate 99.8%
Registries covered 6
Cost for pre-install check Free

Frequently Asked Questions

Is youtube-summarize on ClawHub safe? This skill name was identified in the ClawHavoc campaign as a typosquat used to distribute malware. Check the SkillShield directory for current status and scan results.

How do I know if a skill I already installed is malicious? Run skillshield scan ./SKILL.md against the skill's definition file. If you've already run a skill and suspect compromise, rotate your API keys, SSH keys, and any credentials stored on the machine. Check ~/.ssh/, your .env files, and browser saved passwords.

Does SkillShield replace ClawSecure? No. ClawSecure provides OWASP-based compliance scoring for architecture-level risk. SkillShield provides pre-install malware detection and supply chain scanning. They address different layers of the security stack.

Is SkillShield free? The directory browse and MCP scanner are free. The CLI (npm install -g skillshield) is free. No signup required for any of these.

What about skills from registries other than ClawHub? SkillShield scans six registries: ClawHub, SkillsMP, Skills.lc, MCP Registry, MCPMarket, and Awesome MCP. The CLI can also scan any local SKILL.md file regardless of source.


Scan Now

The ClawHavoc campaign is active. CVE-2026-25253 exposed thousands of instances. The supply chain threat to OpenClaw developers is documented and ongoing.

Check your skills before they check your credentials.


Sources


Catch risky skills before they run.

SkillShield scans skills, MCP servers, and prompt-bearing tool surfaces before they reach production.

Get early access