CloudSecurityPartners-skills

skill-security

Security toolkit for Claude Code — audit skills/plugins for prompt injection and backdoors, or conduct full codebase security reviews with an agent team.

Skills

Skill Security Audit (/skill-audit)

Performs a six-phase security audit of any Claude Code skill or plugin:

1. Target Acquisition — clone repo or verify local path, build file manifest, query OpenSSF Security Scorecard
2. Permissions Analysis — assess allowed-tools against purpose, flag dangerous combinations
3. Prompt Injection Analysis — detect override attempts, concealment, authority manipulation, encoded content
4. Code Execution Analysis — scan scripts and inline code for network ops, sensitive file access, obfuscation
5. Hook Analysis — highest priority — hooks execute without user consent
6. Behavioral Simulation — reason about what the skill actually does when loaded

Produces a structured report with findings by severity and a verdict: SAFE, CAUTION, or DO NOT INSTALL.

Security Review (/security-review)

Conducts a full codebase security review using an agent team:

1. Project Analysis — architecture briefing for the review team
2. Tool Scanning — semgrep, trufflehog, and trivy (deterministic tools)
3. Expert Triage — four parallel agents: SAST triage, dependency triage, targeted expert, broad expert
4. Report Writing — consolidated draft with deduplicated, severity-rated findings
5. Round Table — consensus-driven debate where agents challenge each other's findings

Produces a final report at security-review/report-final.md with confirmed findings, severity ratings, and an appendix of uncertain items.

Prerequisites: semgrep, trufflehog, and trivy must be installed on the host machine.

Installation

From the marketplace

1. Add the marketplace:

/plugin marketplace add CloudSecurityPartners/skills

2. Install the plugin:

/plugin install skill-security@CloudSecurityPartners-skills

Or use /plugin interactively and browse the Discover tab.

As a personal skill

Clone the repo and symlink into your Claude skills directory:

git clone https://github.com/CloudSecurityPartners/skills.git
ln -s "$(pwd)/skills/skills/skill-security-audit" ~/.claude/skills/skill-security-audit
ln -s "$(pwd)/skills/skills/security-review" ~/.claude/skills/security-review

Usage

/skill-audit https://github.com/someone/their-skill
/skill-audit /path/to/local/skill

/security-review
/security-review /path/to/project

Threat categories (skill audit)

ID	Category	Examples
PI	Prompt Injection	Override instructions, concealment, authority manipulation
CE	Code Execution	Shell commands, eval/exec, decode-then-execute chains
DE	Data Exfiltration	curl/wget to external URLs, DNS exfil, env harvesting
SC	Supply Chain	Modifying other plugins, settings, CLAUDE.md, weak repo scorecard
BD	Backdoor	Obfuscated payloads, base64, zero-width characters
PE	Permission Escalation	Overly broad allowed-tools, dangerous tool combos