Skip to main content

ClawHub Skills: What My GitHub Mirror Scan Found

My scan of a public ClawHub GitHub mirror reported historical security findings. The sample is not today's marketplace catalog.

FHFinn Hillebrandt
AI Tools
ClawHub Skills: What My GitHub Mirror Scan Found
Links marked with * are affiliate links. If a purchase is made through such links, we receive a commission.

You found a ClawHub skill that fits your workflow. But do you know which instructions and scripts you are bringing onto your machine?

ClawHub is the public registry for OpenClaw skills. A skill is a versioned bundle with a SKILL.md and possible supporting files. The ClawHub documentation describes automated audits of submitted releases.

That made me curious.

How safe are these skills? In the analysis published on April 6, 2026, I scanned 16,797 entries from the public GitHub mirror with a rule-based scanner. My results list at least one remaining hit for 8,129 entries.

Important note upfront.

Each hit needs manual review to distinguish a possible risk from a false positive.

Before installing a skill, check the audit status for the specific release, its requested permissions, and the files it contains.

TL;DRKey Takeaways
  • ClawHub is the public registry for OpenClaw skills, which are versioned bundles with a SKILL.md.
  • My scan covered 16,797 mirror entries, with at least one pattern hit in 8,129.
  • The GitHub mirror was a snapshot, not a view of today's ClawHub releases.

1. What is the ClawHub?

The ClawHub is the official marketplace for OpenClaw. Skills are to the ClawHub what apps are to the App Store. Small packages that add new capabilities to OpenClaw. That might be a PDF parser, a database connector, a wrapper around an external API, or a full workflow.

Installation is dead simple. One command in the terminal, and the skill is ready to go. That's exactly what makes the ClawHub so attractive. And at the same time so risky.

ClawHub's documentation describes audits for submitted release artifacts and their audit status.

2. The Methodology Behind the Analysis

I used the public GitHub mirror (clawhub-skills-repo) rather than downloading the entire ClawHub marketplace. I downloaded around 168 skills directly from the marketplace; other attempts returned HTTP 429 errors.

The analysis was published on April 6, 2026, but I cannot document its exact collection date. The raw data, scanner rules, and refinement filter are no longer available, so the individual figures cannot be reproduced.

I successfully scanned 16,797 mirror entries. The mirror was a time-bound snapshot. Skills removed later could still be included, while newer ones were missing.

The scanner itself is a deterministic tool. That means it works with fixed rules (regex and heuristics), no LLM, no probabilities. Every line of code and every file gets checked against a list of known risk patterns. When the scanner finds a match, it logs a finding, categorizes it, and assigns a severity level.

The categories are:

  • Supply Chain: unpinned pip/npm installs, curl-to-bash, package tampering
  • Data Exfiltration: SSH key accesses, hardcoded endpoints, browser data reads
  • Privilege Escalation: sudo, capability abuse, root execution
  • Persistence: cron jobs, profile modifications, CLAUDE.md writes
  • Destructive Patterns: rm -rf, shutil.rmtree, mass deletions
  • Code Execution: eval, shell pipes, reverse shells, unsafe deserialization
  • Hardcoded Secrets: embedded API keys, passwords, tokens
  • Prompt Injection: attempts to override roles or safety instructions
  • Obfuscation: homoglyphs, unicode tricks, unusual encodings
  • SSRF, Sandbox Escape, Suspicious Files, Resource Exhaustion as additional categories

I also used a refinement filter for three patterns that commonly create false positives: Markdown code blocks, homoglyphs in multilingual documentation, and HTTP examples in .md files. In my results, it removed 261,451 instances.

3. The Results at a Glance

The chart shows how the 16,797 mirror entries in my analysis split between findings and no findings.

No findings
51.6%
With security finding
48.4%
No findings: 8,668 skills
With security finding: 8,129 skills
Source: own mirror scan, Apr 6, 2026
gradually.ai

At least one security finding remained for 8,129 of 16,797 entries, or 48.4% of this mirror sample.

In the remaining 51.6%, the scanner found no matching pattern under its rules after filtering.

My analysis counted 182,258 individual hit instances. One skill can contribute several hits, which is why some publishers can have many instances despite publishing only a few skills.

The scanner classified 26,807 hits as Critical, or 14.7% of all instances.

4. Severity Breakdown

Severity levels for the scanner hits:

Critical
26,807
High
139,664
Medium
15,787
Source: own mirror scan, Apr 6, 2026
gradually.ai

High accounted for 76.6% of the hits. Some involved patterns such as homoglyphs and unpinned installs.

Critical accounted for 14.7%, or 26,807 hit instances. Examples included hardcoded API keys, embedded passwords, shell pipes with eval, and executing third-party scripts through curl and bash.

Medium accounted for 8.7% of hits, including some file access, configuration changes, and permission requests.

5. The Threat Categories

I grouped the scanner hits into 13 categories. The charts show both the share of affected skills and the share of all hit instances.

5.1. By Share of Affected Skills

The first chart shows the share of 16,797 mirror entries with at least one hit in each category:

Supply Chain
21.4%
Data Exfiltration
20.1%
Privilege Escalation
12.1%
Persistence
9.7%
Destructive Patterns
9.0%
Code Execution
8.8%
Obfuscation
6.8%
Hardcoded Secrets
4.1%
Prompt Injection
3.0%
SSRF
2.4%
Sandbox Escape
1.5%
Suspicious Files
0.6%
Resource Exhaustion
0.3%
Source: own mirror scan, Apr 6, 2026
gradually.ai

Supply Chain was the largest category at 21.4%. Hits included unpinned installs, curl-to-bash, and package tampering.

Data Exfiltration followed at 20.1%. The patterns included access to sensitive directories, hardcoded endpoints, and unexplained network operations.

Privilege Escalation followed at 12.1%, Persistence at 9.7%, and Destructive Patterns at 9.0%.

5.2. By Share of All Hit Instances

The second chart groups hit instances by category. A skill may contribute several hits:

Obfuscation
55.6%
Data Exfiltration
17.3%
Supply Chain
7.0%
Privilege Escalation
5.9%
Persistence
3.9%
Destructive Patterns
3.4%
Code Execution
3.4%
Hardcoded Secrets
1.3%
Prompt Injection
0.7%
SSRF
0.7%
Sandbox Escape
0.6%
Suspicious Files
0.1%
Resource Exhaustion
0.0%
Source: own mirror scan, Apr 6, 2026
gradually.ai

Obfuscation made up 55.6% of hit instances. I found at least one pattern in this category in 6.8% of mirror entries. Homoglyphs, Unicode tricks, and unusual encodings can reflect intentional concealment or multilingual documentation; each case needs manual review.

By hit instances, Data Exfiltration followed at 17.3%, Supply Chain at 7.0%, Privilege Escalation at 5.9%, Persistence at 3.9%, and Destructive Patterns at 3.4%.

6. What the Most Important Categories Actually Mean

The examples below explain the patterns behind the scanner categories.

6.1. Supply Chain (21.4%)

The scanner flags install instructions that skip version pinning. Commands like pip install requests instead of pip install requests==2.31.0. Or shell pipes like curl https://example.com/install.sh | bash. These commands are normal in the open source world, but they have a catch. If the package or script gets compromised later, anyone installing it fresh gets compromised too.

6.2. Data Exfiltration (20.1%)

This is about patterns that look like data being sent out. Accesses to ~/.ssh/, reading .env files, hardcoded IP addresses like http://127.0.0.1:8765, or suspicious HTTP requests to endpoints outside the official providers. After the refinement filter, what's left is mostly hits outside Markdown code blocks.

6.3. Privilege Escalation (12.1%)

Every time a script uses sudo, requests root privileges, or changes system configurations, it lands in this category. For system tools, that's completely normal. A skill that installs packages often needs sudo. A skill that just parses PDFs doesn't. The scanner can't tell the difference. You can, if you look at the code.

6.4. Persistence (9.7%)

Persistence means a skill embeds itself in the system long-term. Classic patterns are cron jobs, modifications to shell profiles (.bashrc, .zshrc), or new entries in autostart folders. Writes to CLAUDE.md files also fall into this category. Persistence isn't inherently bad, but it should always be transparent.

6.5. Destructive Patterns (9.0%)

Mass deletions with rm -rf, shutil.rmtree, or similar commands. Sometimes that's legitimate (a cleanup skill has to be able to delete things). Sometimes it's an accident waiting to happen. And in the worst case, it's intentional.

6.6. Code Execution (8.8%)

Unsafe execution of third-party code. eval(), exec(), unsafe deserialization with pickle, reverse shells, shell pipes with user input. The scanner knows the typical patterns and flags them. When in doubt, a hit here means you should stay away until you've actually read the code.

7. Publishers with the Most Hits in the Mirror Scan

In my mirror sample, publisher pepe276 had 37,978 hits. That's 20.8% of the hits in this sample. The table lists the 20 publishers with the most hit instances:

pepe276
Total findings
37,978
Critical
19
High
37,948
Medium
11
Scanner signal
Critical hits
cooperun
Total findings
5,201
Critical
4,982
High
192
Medium
27
Scanner signal
Many critical hits
yuangu260
Total findings
4,582
Critical
26
High
4,550
Medium
6
Scanner signal
Critical hits
jimliu
Total findings
4,575
Critical
6
High
4,555
Medium
14
Scanner signal
Critical hits
ciklopentan
Total findings
4,235
Critical
2
High
4,233
Medium
0
Scanner signal
Critical hits
keenone
Total findings
2,681
Critical
3
High
2,677
Medium
1
Scanner signal
Critical hits
horosheff
Total findings
2,441
Critical
0
High
2,440
Medium
1
Scanner signal
Mostly high hits
qiumr
Total findings
2,421
Critical
6
High
2,412
Medium
3
Scanner signal
Critical hits
snail3d
Total findings
2,149
Critical
363
High
1,650
Medium
136
Scanner signal
Many critical hits
yoborlon-alpha
Total findings
2,059
Critical
1
High
2,058
Medium
0
Scanner signal
Critical hits
deerleo
Total findings
1,687
Critical
8
High
1,675
Medium
4
Scanner signal
Critical hits
keeper1978
Total findings
1,591
Critical
0
High
1,591
Medium
0
Scanner signal
Mostly high hits
mirra87654321
Total findings
1,319
Critical
1
High
1,318
Medium
0
Scanner signal
Critical hits
satoshistackalotto
Total findings
1,250
Critical
0
High
1,250
Medium
0
Scanner signal
Mostly high hits
rsvbitrix
Total findings
1,230
Critical
1
High
1,229
Medium
0
Scanner signal
Critical hits
h8kxrfp68z-lgtm
Total findings
1,161
Critical
6
High
1,154
Medium
1
Scanner signal
Critical hits
mixx85
Total findings
1,041
Critical
25
High
1,012
Medium
4
Scanner signal
Critical hits
chorus12
Total findings
1,008
Critical
0
High
1,008
Medium
0
Scanner signal
Mostly high hits
s7cret
Total findings
950
Critical
0
High
938
Medium
12
Scanner signal
Mostly high hits
offflinerpsy
Total findings
941
Critical
1
High
938
Medium
2
Scanner signal
Critical hits

The full aggregation covers 50 publishers. High hit counts can result from extensive documentation, many files, or actual risk patterns, and need manual review.

Almost all pepe276 hits were classified as High. That can result from many skills or extensive files.

For cooperun, the scanner classified 4,982 hits as Critical. This is a reason to inspect the specific files and findings manually.

8. Which Patterns Did the Refinement Filter Remove?

I used a refinement filter to reduce expected false positives from three pattern groups:

Homoglyphs (HG-001)
160,889
HTTP examples (SC-004)
54,034
Code in Markdown fences
46,528
Source: own mirror scan, Apr 6, 2026
gradually.ai

The filter removed 261,451 instances in my analysis, more than the 182,258 hits that remained after filtering.

Homoglyphs in Markdown accounted for the largest group, with 160,889 removed hits. Characters in Chinese, Russian, or Arabic text may look like obfuscation to a scanner even when they are part of multilingual documentation.

HTTP examples in documentation accounted for 54,034 removed hits. Pattern SC-004 covered API examples in .md files.

Code in Markdown fences accounted for 46,528 removed hits. These included examples such as curl -X POST ... or pip install tensorflow.

A filter can also miss real risks inside code blocks. Malicious code in a Markdown block would no longer be flagged after those hits were removed.

9. What This Means for You

No need to panic. But no blind trust either. The numbers are a wake-up call, not the end of the world. Here are my concrete recommendations for working with ClawHub skills:

  • Check the publisher. A verified status is not a security guarantee. Consider the publisher's reputation, the specific version, and its changelog instead of treating a status alone as proof of trust.
  • Read the SKILL.md. If you really need a skill, open the SKILL.md and the scripts it references. Look for sudo commands, curl installs, accesses to .ssh or .env. If you don't understand some of the code, ask a colleague or have Claude or ChatGPT explain it.
  • Use a sandbox. OpenClaw supports restricted permissions. Give a skill only the rights it actually needs. No root rights for a PDF parser. No network access for a file renamer.
  • Keep OpenClaw up to date. New releases may include security fixes. Check the release notes for the version you use.
  • Avoid sensitive data. Do not run an unreviewed skill with production credentials, payment-provider API keys, or personal health data. Check which data and services it can access first.

If security matters more to you than feature breadth, take a look at the secure OpenClaw alternatives. Projects like OpenFang and IronClaw are built with security-first principles from day one. The selection of skills is much smaller, but the quality control is stricter.

10. Additional Security Research on Agent Skills

Snyk's ToxicSkills study reports on a sample of 3,984 skills. The researchers distinguish between security findings and confirmed malicious payloads:

  • 1,467 skills (36.82%) had at least one security finding
  • 534 skills (13.4%) had at least one finding classified as critical
  • 76 malicious payloads were confirmed
  • 8 of the confirmed malicious skills were still available on ClawHub when the study was published

Adversa also published a hardening guide in the OpenClaw Security Guide 2026. OpenClaw says CVE-2026-25253 (one-click RCE via a manipulated Control UI parameter, CVSS 8.8) was patched in version 2026.1.29.

What remains.

An automated audit status does not replace reviewing the specific release. Read the SKILL.md, compare requested permissions with the stated purpose, and inspect referenced files before installing. If access is unclear, test the skill in isolation without real credentials.

Frequently Asked Questions

ClawHub is the public registry for OpenClaw skills. A skill is a versioned text bundle with a SKILL.md and possible supporting files. OpenClaw can install and update ClawHub skills in the active workspace.

ClawHub does not give a total. Its public security documentation shows audit status and findings only for individual releases. A release marked “Malicious” should not be installed, according to ClawHub. Other findings are review signals, not proof of malware.

Not automatically. ClawHub audits releases and shows status and findings, but a “Pass” does not replace your own review. Do not install a release marked “Malicious”.

“Critical” is only a severity for individual findings, while “Medium” and “High” also exist as a release's risk level. The risk level describes the potential blast radius of the requested authority: “Medium” means meaningful authority such as account access or data changes, “High” high-impact authority, severe findings, or malicious signals. Individual findings range from Info to Critical, and higher severities weigh more in risk level and audit status. ClawHub does not publish exact thresholds, so a label alone is not a verdict on the whole skill.

Read the audit status, risk level, and findings, and compare requested permissions, credentials, and environment variables with the skill's stated purpose. Open its SKILL.md and referenced files and scripts, and check the owner, source, and version. A “Pass” status does not replace your judgment. If you cannot explain what a skill runs, test it in isolation without real credentials.

By hit instances, obfuscation: it made up 55.6% of hits, and I found at least one pattern of this kind in 6.8% of mirror entries. Some of those can be homoglyphs in multilingual documentation. By affected skills, Supply Chain was the largest category at 21.4%. My mirror scan counted individual pattern hits, so one file could contribute several findings and one skill could include multiple affected files. A share of hit instances is not a share of distinct skills. ClawHub's current audits assess individual releases instead.
FH

Finn Hillebrandt

AI Expert & Blogger

Finn Hillebrandt is the founder of Gradually AI, an SEO and AI expert. He helps online entrepreneurs simplify and automate their processes and marketing with AI. Finn shares his knowledge here on the blog in 50+ articles as well as through the AI Business Club.

Learn more about Finn and the team, follow Finn on LinkedIn, join his Facebook group for ChatGPT, OpenAI & AI Tools or do like 17,500+ others and subscribe to his AI Newsletter with tips, news and offers about AI tools and online business. Also visit his other blog, Blogmojo, which is about WordPress, blogging and SEO.