Skip to content

Your exploit code is being auto-distributed as Claude Code skills — malware alerts on end-user machines #4

Closed

Description

@TheSoulGiver

Notification

Your repository's content (offensive-sqli/SKILL.md) has been automatically ingested by majiayu000/claude-skill-registry and redistributed as a Claude Code skill to end users.

Detection: Windows Defender flags it as Backdoor:PHP/Perhetshell.A!dha (Severity: Severe)

This is causing repeated malware quarantine alerts on machines that clone the registry.

See full audit: majiayu000/claude-skill-registry#27

Activity

  1. TheSoulGiver commented on May 3, 2026

    @TheSoulGiver
    Author

    Full Malware Audit Report

    This repository's content was detected as malware by Windows Defender on an end-user machine. This comment serves as a formal record of the full audit.

    30 malicious files across 9 source repos were distributed via majiayu000/claude-skill-registry

    Malware Signatures (all Severity: SEVERE):

    Signature Type
    Backdoor:PHP/Perhetshell.A!dha PHP backdoor
    Backdoor:PHP/Perhetshell.B!dha PHP backdoor
    Backdoor:Perl/TurtlePerlsh.A!dha Perl backdoor
    Backdoor:PHP/Remoteshell.F PHP remote shell
    Trojan:PowerShell/ReverseShell.HNAA!MTB PowerShell reverse shell
    Trojan:PowerShell/Openclaw.GVB!MTB PowerShell trojan
    Trojan:AIAgent/ClawHavoc.GVD!MTB AI agent trojan
    Trojan:AIAgent/ClawHavoc.A!AMTB AI agent trojan

    Malicious source repos:

    1. blacklanternsecurity/red-run — 7 skills (PHP backdoors)
    2. trilwu/secskills — 5 skills (PHP backdoors)
    3. macaugh/super-rouge-hunter-skills — 4 skills (PHP backdoors)
    4. zebbern/SecOps-CLI-Guides — 4 skills (PowerShell reverse shells)
    5. AgentSecOps/SecOpsAgentKit — 4 skills (PowerShell reverse shells)
    6. openclaw/skills — 3 skills (trojans, repo now 404)
    7. YPYT1/All-skills — 2 skills (trojans, repo now 404)
    8. SnailSploit/Claude-Red — 1 skill (PHP backdoor)
    9. aifinlab/FinClaw — 1 skill (AI agent trojan)

    All files disguised as Markdown (SKILL.md) but contain real exploit code. This is a supply chain attack on the Claude Code skill ecosystem.

    This has been reported to GitHub Trust & Safety. Full technical audit: majiayu000/claude-skill-registry#27

  2. SnailSploit commented on May 6, 2026

    @SnailSploit
    Owner

    Maintainer of SnailSploit/Claude-Red here.

    This is a false positive. My repo should be removed from this list.

    No executable code exists in this repo

    Claude-Red contains 38 SKILL.md files (markdown) and one Python converter script. Zero PHP. Zero Perl. Zero PowerShell. Zero compiled binaries. You can verify this in seconds:

    git clone https://github.com/SnailSploit/Claude-Red && find Claude-Red -type f | grep -vE '\.(md|py)$'

    The only non-markdown, non-Python files are LICENSE and README.md.

    What Defender actually matched

    The Backdoor:PHP/Perhetshell.A!dha signature triggered on a literal payload string inside offensive-sqli/SKILL.md — a markdown file documenting SQL injection techniques:

    ' UNION SELECT '<?php system($_GET["cmd"]); ?>' INTO OUTFILE '/var/www/html/shell.php' --
    

    This is a SQLi-to-webshell example. The same string appears in the OWASP Testing Guide, PayloadsAllTheThings, HackTricks, and every SQLi reference on the internet. Defender does content-based signature matching regardless of file extension — it will flag a .txt, .md, or .pdf containing PHP shell strings the same way it flags an actual .php file. That's a known limitation of signature-based detection and exactly why file type context matters.

    Your report's core claim is wrong for this repo

    You wrote: "All files disguised as Markdown (SKILL.md) but contain real exploit code."

    That may be true for the other 8 repos on this list — the ClawHavoc/Openclaw campaign is well-documented (Snyk ToxicSkills, Repello, Mobb). But Claude-Red is not part of that campaign. My files are not "disguised as Markdown" — they are markdown. They document offensive methodology the same way Metasploit documents exploit modules or SecLists documents payload patterns. Documenting a technique is not the same as deploying it.

    Context

    Claude-Red is a published offensive security skill library (1.2k+ stars, MIT licensed). I'm an independent security researcher with 22+ published CVEs, a Linux kernel contributor, Hakin9 author, and creator of the AATMF framework. This is legitimate security tooling.

    The actual supply chain issue

    claude-skill-registry scraped and redistributed my repo's content without authorization or review. If you're auditing supply chain integrity, that's the vector — not the upstream source repos whose content was ingested without consent.

    Please update the report to remove SnailSploit/Claude-Red from the malicious repo list.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions