⚠ 341 malicious skills found on ClawHub this month

Your OpenClaw is wide open.

Scan your skills for malware. Monitor your instance 24/7. Harden your config in one click. The security layer OpenClaw doesn't have — yet.

Scan Free Now View Plans
Terminal — clawcop scan
Example output
$ npx clawcop scan --instance ~/. openclaw # Scanning 47 installed skills... # Checking config, ports, credentials, permissions...   ✗ CRITICAL rss-reader — data exfiltration to external server ✗ CRITICAL smart-mailer — prompt injection bypasses safety ✗ HIGH sandbox mode disabled — shell runs on host ✗ HIGH 3 API keys in plaintext .env ⚠ MEDIUM DM policy set to "open" — anyone can message ✓ PASS gateway not exposed to public internet   Found 2 critical, 2 high, 1 medium issue. → Fix all automatically with Clawcop Shield: clawcop.dev/fix
341
Malicious skills on ClawHub
Koi Security, Feb 2026
42,665
Exposed instances on internet
Shodan / Dvuln scan
93.4%
Vulnerable to auth bypass
AdwaitX Research, 2026
26%
Of all skills contain vulnerabilities
Cisco AI Defense

You gave an AI agent the keys
to your entire digital life.

Your email. Your files. Your API keys. Your shell. OpenClaw is incredible — but without security, you're running the most powerful attack surface ever put on a consumer machine.

🧬

Skill Supply Chain Attacks

ClawHub has no vetting process. The #1 ranked skill was literal malware — silently exfiltrating your data while you slept. 341 malicious packages found in a single week.

ClawHavoc campaign deployed infostealers, keyloggers, and backdoors across hundreds of skills — Koi Security
💉

Prompt Injection via Chat

A single crafted message in a WhatsApp group, a Discord DM, or a forwarded email can hijack your agent. It'll dump your SSH keys and you won't even see it happen.

A tester sent one email — the agent leaked private keys from the machine without any confirmation — Kaspersky
🔓

Credentials in Plaintext

API keys, session tokens, OAuth credentials — stored in Markdown files and .env files in the agent's reachable filesystem. 1,800+ instances actively leaking credentials on the open internet.

CVE-2026-25253: Critical RCE (CVSS 8.8) — steal auth tokens via a single malicious link
🧠

Memory Poisoning

OpenClaw stores web scrapes, messages, and skill outputs in the same memory — without trust levels. An attacker who poisons your agent's memory controls its behavior for weeks.

Modifiable memories and system prompts persist into future chat sessions — HiddenLayer
🌐

Agent-to-Agent Contagion

Moltbook — the AI social network — has 37,000 agents processing untrusted content from other agents. Prompt injection payloads spread virally between instances.

Moltbook created a laboratory where agents constantly process untrusted data — Zenity Labs
👻

No Approval Required

By default, OpenClaw executes tool calls without explicit user approval. Shell commands, file writes, network requests — the LLM decides, and the LLM can be deceived.

The LLM cannot distinguish between trusted instructions and untrusted data — it becomes a "confused deputy" — Cisco
"From a capability perspective, OpenClaw is groundbreaking. From a security perspective, it's an absolute nightmare."
Cisco AI Defense — Security Research Team
"If you can't understand how to run a command line, this is far too dangerous of a project for you to use safely."
Shadow — OpenClaw Core Maintainer
"AI agents get credentials to your entire digital life. Unlike browser extensions in a sandbox, these agents operate with full privileges."
Ian Ahl — Permiso Security

Three commands. Five minutes.
Actually secure.

Clawcop works alongside your OpenClaw instance — not against it. Keep using everything you love. Remove everything that can hurt you.

01

Scan your instance

One command scans every installed skill for malware, data exfiltration, and prompt injection. Checks your config for exposed ports, plaintext credentials, and dangerous permissions. Takes under 60 seconds.

$ npx clawcop scan
02

Fix with one click

Every finding comes with a severity rating, an explanation a human can understand, and a one-click fix. Remove the malicious skill. Rotate the exposed key. Enable sandbox mode. Done.

$ npx clawcop fix --all
03

Monitor 24/7

Upgrade to Shield and Clawcop watches your instance continuously. New skill installs get scanned automatically. Config drift triggers alerts. New CVEs get flagged before they're exploited. You sleep.

→ clawcop.dev/dashboard

Everything between your OpenClaw
and the threat actors.

🔍

Skill Scanner

Multi-layer analysis of every ClawHub skill: static code analysis, behavioral dataflow tracing, LLM semantic analysis for covert prompt injection, and VirusTotal cross-reference. Goes beyond one-time scanning with hardening + monitoring.

Free
🛡

Instance Hardening

Audits your OpenClaw config against 40+ security checks: sandbox mode, DM policies, tool permissions, credential storage, port exposure, exec approvals, and network isolation. One-click remediation for every finding.

Free
📡

Continuous Monitoring

Real-time surveillance of your running instance. Alerts on new skill installs, configuration changes, unexpected network calls, credential access, and anomalous agent behavior. Delivered via the same channels you use with OpenClaw.

Shield
🚨

Threat Intelligence

Weekly ClawHub security bulletins. Real-time alerts when new malicious skills are detected. CVE tracking for OpenClaw core. Curated, prioritized, actionable — not a firehose of noise.

Shield
🧪

Prompt Injection Shield

Analyzes inbound messages across all connected channels for adversarial patterns before they reach your agent. Blocks injection attempts from emails, DMs, web content, and Moltbook interactions.

Shield
🏢

Fleet Management

Multi-instance dashboard for teams and enterprises. Enforce skill allowlists across all instances. Centralized policy management. Audit logging. SSO/SAML. Compliance reporting.

Enterprise

Less than your monthly API costs.
More than worth it.

You're spending $30-70/month on LLM APIs to power your OpenClaw. Spend a fraction of that to make sure it isn't robbing you blind.

Scanner
$ 0
Free forever. Open source.
One-time scan of your skills and config. Catch the obvious stuff. A good start — but threats don't stop after one scan.
  • Scan all installed skills for malware
  • Detect prompt injection & data exfil
  • Config audit (40+ security checks)
  • Severity-rated report with fix guidance
  • Community detection rules
Run Free Scan
Armor
$ 29 /mo
$279/year (save 20%) · up to 5 instances
For teams and power users running multiple instances. Fleet dashboard, advanced threat intelligence, and the full security stack.
  • Everything in Shield
  • Up to 5 OpenClaw instances
  • Team dashboard with role management
  • Skill allowlist/denylist enforcement
  • Advanced threat intelligence feed
  • Memory integrity analysis
  • Agent behavior anomaly detection
  • Credential rotation reminders
  • Priority email support
Start Armor — 14 Days Free

Running 10+ instances or need SSO, compliance, and audit logs?
Talk to us about Enterprise →

Questions you're probably asking.

Why not just use Cisco's Skill Scanner?

Cisco's scanner is a great one-time check (we build on their open-source work and credit them). Clawcop adds continuous monitoring, automatic remediation, config hardening, prompt injection detection, and threat intelligence. Cisco's scanner tells you there's a problem. Clawcop fixes it and makes sure it doesn't happen again.

Will this slow down my OpenClaw?

No. The scanner runs as a one-time audit (takes ~60 seconds). Continuous monitoring runs as a lightweight sidecar process — it watches logs and config files, it doesn't intercept your agent's execution. Typical overhead: <2% CPU, <50MB RAM.

Can I trust you with access to my instance?

Clawcop is open-source at its core — inspect every line of code on GitHub. The scanner runs entirely locally. The monitoring dashboard receives security events only (not your messages, files, or data). We process zero PII. We're also working on SOC 2 Type II certification.

I'm technical. Do I really need this?

Shadow — an OpenClaw core maintainer — said the project is "far too dangerous" for non-technical users. Kaspersky found 512 vulnerabilities in a single audit. Even if you're technical, are you auditing every skill's source code? Checking ClawHub for new malicious packages daily? Monitoring your instance's network calls 24/7? That's what we do.

What if OpenClaw fixes its own security?

We hope they do — and we help them (we contribute upstream). But even with perfect defaults, you still need: independent scanning of third-party skills, continuous monitoring for config drift, threat intelligence for new attack patterns, and audit evidence. WordPress is much more secure than it was in 2010. Wordfence still has 5 million active installs.

Is this just for OpenClaw?

Today, yes. Our scanner and monitoring are purpose-built for OpenClaw's architecture. But the attack surface — autonomous agents with system access and community plugins — is identical across Copilot, Gemini agents, and every MCP-based system. We're starting where the need is most acute and expanding from there.

How is this different from antivirus?

Traditional antivirus catches known malware signatures. OpenClaw's threats are different: prompt injection hidden in natural language, behavioral data exfiltration that looks like normal API calls, and memory poisoning that corrupts the agent over time. We use multi-layer analysis — static, behavioral, LLM semantic, and runtime sandboxing — designed specifically for AI agent threats.

Do you have a refund policy?

14-day free trial on all paid plans. No credit card required to start. If you're not happy after that, email us and we'll refund the current month, no questions asked. We'd rather have happy users than locked-in users.

Your OpenClaw runs 24/7.
So should your security.

Free scan takes 60 seconds. No signup required. See exactly what's exposed — then decide.

$ npx clawcop scan