Gemini CLI Security Skill

Overview

This skill adapts the Gemini CLI Security Extension's analysis methodology for the agent-studio framework. The original extension uses two MCP server patterns — a security analysis server and an OSV-Scanner integration — to provide dual-vector coverage. This skill implements equivalent analysis using native Claude Code tools (WebFetch for OSV.dev API, Grep/Bash for static analysis patterns).

Source repository: https://github.com/gemini-cli-extensions/security License: Apache 2.0 Performance: 90% precision, 93% recall (OpenSSF CVE benchmark, TypeScript/JavaScript)

When to Use

• Before merging pull requests to detect introduced vulnerabilities
• During security reviews of new code changes
• For dependency auditing against known CVE databases
• For LLM-integrated applications requiring prompt injection defense review
• As part of CI/CD pipeline security gates

Iron Law

NO PRODUCTION CODE WITHOUT SECURITY ANALYSIS FOR AUTH/SECRETS/EXTERNAL-INPUT HANDLERS

All code paths handling authentication, hardcoded values, external input, or AI model outputs MUST be analyzed before production deployment.

Vulnerability Coverage

Category 1: Secrets Management

Pattern	Detection Method
Hardcoded API keys	Grep for key patterns + entropy analysis
Hardcoded passwords	Credential keyword detection
Private keys in source	PEM block / base64 key detection
Encryption keys	Symmetric key constant patterns

Category 2: Injection Attacks

Attack Type	Examples
SQL injection	String concatenation in queries
XSS	Unescaped user content in HTML/JS output
Command injection	Shell exec with user-controlled args
SSRF	User-controlled URLs in server requests
Template injection	Unsanitized user input in template engines

Category 3: Authentication Flaws

Flaw	Detection
Session bypass	Missing auth middleware
Weak tokens	Predictable token generation
Insecure password reset	Token-less or email-only resets
Missing MFA enforcement	Auth flows without 2FA checks

Category 4: Data Handling

Issue	Detection
Weak cryptography	MD5/SHA1 for secrets; DES/RC4 usage
Sensitive data in logs	PII/credential patterns in log statements
PII violations	Unencrypted PII storage or transmission
Insecure deserialization	Unsafe pickle/eval/deserialize calls

Category 5: LLM Safety (Novel)

Risk	Detection
Prompt injection	User content injected into LLM prompts without sanitization
Unsafe output handling	LLM output used in exec/eval/shell without validation
Insecure tool integration	Tool calls with unchecked LLM-provided parameters

Usage

Invocation

// From an agent
Skill({ skill: 'gemini-cli-security' });

// With arguments via Bash integration
Skill({ skill: 'gemini-cli-security', args: 'src/ --scan-deps' });

Workflow Execution

# Analyze code in a directory
node .claude/skills/gemini-cli-security/scripts/main.cjs --target src/

# Scan dependencies for CVEs
node .claude/skills/gemini-cli-security/scripts/main.cjs --scan-deps

# JSON output for CI integration
node .claude/skills/gemini-cli-security/scripts/main.cjs --target . --json

# Scoped analysis with natural language
node .claude/skills/gemini-cli-security/scripts/main.cjs --target src/auth/ --scope "focus on token handling and session management"

Output Format

Default output (markdown report):

## Security Analysis Report

### CRITICAL

- [AUTH-001] Hardcoded API key found in src/config.ts:42
  Pattern: `const API_KEY = "sk-..."`
  Remediation: Move to environment variable

### HIGH

- [INJ-002] SQL injection risk in src/db/users.ts:87
  Pattern: String concatenation in query builder
  Remediation: Use parameterized queries

### Dependencies

- lodash@4.17.15 → CVE-2021-23337 (HIGH) - Prototype pollution
  Fix: Upgrade to lodash@4.17.21+

JSON output (--json flag):

{
  "findings": [
    {
      "id": "AUTH-001",
      "severity": "CRITICAL",
      "category": "secrets",
      "file": "src/config.ts",
      "line": 42,
      "description": "Hardcoded API key",
      "remediation": "Move to environment variable"
    }
  ],
  "dependencies": [
    {
      "package": "lodash",
      "version": "4.17.15",
      "cve": "CVE-2021-23337",
      "severity": "HIGH",
      "fix": "4.17.21"
    }
  ],
  "summary": {
    "critical": 1,
    "high": 2,
    "medium": 3,
    "low": 0,
    "precision": 0.9,
    "recall": 0.93
  }
}

OSV.dev Dependency Scanning

The skill integrates with the OSV.dev API (no authentication required) to check dependencies:

// OSV.dev batch query endpoint
WebFetch({
  url: 'https://api.osv.dev/v1/querybatch',
  prompt: 'Extract vulnerability IDs, severity, and affected versions for these packages',
});

Supported ecosystems: npm, PyPI, RubyGems, Maven, Go, Cargo, NuGet, Packagist

GitHub Actions Integration

The original extension supports PR analysis via GitHub Actions. This skill includes an equivalent workflow template:

# .github/workflows/security.yml
name: Security Analysis
on: [pull_request]
jobs:
  security:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run security analysis
        run: node .claude/skills/gemini-cli-security/scripts/main.cjs --target . --json

Implementation Notes

Why native tools over MCP servers: The original extension uses two MCP servers (security analysis server + OSV-Scanner binary). This skill uses native Claude Code tools instead:

• WebFetch replaces OSV-Scanner for dependency CVE lookups (OSV.dev has a public REST API)
• Grep/Bash replace the security analysis server for pattern-based detection
• This approach works immediately without binary installation or session restart

Deviation from source: The original uses Gemini AI for code analysis; this skill uses the pattern-based detection methodology documented in the extension's benchmarking. The AI analysis component can be provided by the invoking agent (security-architect) rather than an embedded AI call.

Assigned Agents

Agent	Role
security-architect	Primary: comprehensive security audits
developer	Supporting: pre-commit security checks
code-reviewer	Supporting: PR review security layer

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New vulnerability pattern found -> .claude/context/memory/learnings.md
• Issue with scanning -> .claude/context/memory/issues.md
• Decision about scope -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.