AI Security Arsenal

A collection of skills for automated security testing, penetration testing, and security assessment workflows with coding agents like Claude Code, Codex, OpenCode, and others.

Each skill is structured as a thin router (SKILL.md, 150-250 lines) with lazy-loaded sibling content (workflows/, methodology/, payloads/, references/, examples/, schemas/). This lets agents load only what a sub-task needs instead of parsing a monolithic document every turn — leaving more context for the actual target code, traffic, and tool output.

Last validated: 2026-04 against frontier coding agents (Claude Opus 4.x, GPT-5.x class).

Skills Overview

---

Skill Architecture

Every skill follows the same layout so agents can navigate them predictably:

skills//
├── SKILL.md           # Thin router (150-250 lines): triggers, decision tree, indexes
├── workflows/         # Step-by-step runbooks loaded per task
├── methodology/       # Deep methodology docs per phase
├── payloads/          # Payload/injection/query files loaded on demand
├── references/        # Static reference material (OWASP, CWE, tool docs)
├── examples/          # Concise tool-call blueprints and worked examples
├── templates/         # Report/finding templates
├── schemas/
│   └── finding.json   # JSON Schema for structured findings
└── scripts/           # Executable helpers (where present)

Every SKILL.md includes:

• When NOT to use — explicit pointers to sibling skills for overlap cases
• Decision tree — how to pick the right workflow
• Parallelism hints — which operations run concurrently vs sequentially
• Sub-agent delegation — when to spawn parallel sub-agents (e.g., one per domain, one per tool, one per auth context)
• Reasoning budget — where extended thinking pays off vs where to execute without it
• Multimodal hooks — screenshots, diagrams, image-based evidence (where relevant)
• Output schema reference — points at schemas/finding.json

Findings validate against the per-skill JSON Schema, so output can be piped directly into SARIF converters, Jira, DefectDojo, or custom triage tooling.

Headline workflows (new in 2026-04)

---

Mobile Security

Android Pentest

Comprehensive Android mobile application penetration testing with full ADB shell access to rooted devices.

Capabilities:

• Static & dynamic analysis of Android applications
• Runtime manipulation with Frida (SSL pinning bypass, root detection bypass, crypto hooks)
• OWASP MASTG compliance testing
• Data extraction (databases, SharedPreferences, storage)
• Component fuzzing (activities, services, content providers, broadcast receivers)

Trigger phrases: pentest this Android app, bypass SSL pinning, extract data from app, MASTG testing

Tools: ADB, Frida, objection, apktool, jadx, Burp Suite

---

iOS Pentest

iOS mobile application security testing with runtime manipulation and binary analysis.

Capabilities:

• Static analysis with class-dump, otool, Hopper
• Runtime manipulation with Frida and Cycript
• SSL pinning and jailbreak detection bypass
• Keychain extraction and data protection analysis
• OWASP MASTG compliance testing
• IPA decryption and binary patching

Trigger phrases: pentest this iOS app, bypass jailbreak detection, extract keychain, iOS security assessment

Tools: Frida, objection, Cycript, class-dump, otool, Clutch/bfdecrypt

---

Mobile MCP for Device Interaction

Both mobile pentesting skills support Mobile MCP for advanced device, emulator, and simulator interactions:

• Screen capture and UI inspection
• Touch automation and gesture simulation
• App lifecycle management
• Complements Frida/Objection for scenarios requiring direct UI interaction

// Add to ~/.claude/mcp.json
{
  "mcpServers": {
    "mobile-mcp": {
      "command": "npx",
      "args": ["-y", "@anthropic/mobile-mcp", "--android"]  // or "--ios"
    }
  }
}

---

Web Security

DAST Automation

Automated Dynamic Application Security Testing using browser-based scanning and traditional tools.

Capabilities:

• Blackbox and greybox (authenticated) testing
• Intelligent browser-based crawling with JavaScript execution
• Multi-domain parallel scanning
• XSS, CSRF, SQLi, SSRF, IDOR, auth bypass detection
• Integration with Nuclei, ZAP, Burp Suite, SQLMap

Trigger phrases: scan domain for vulnerabilities, run DAST on, automated security scan, web penetration test

Tools: Playwright MCP, Nuclei, OWASP ZAP, Burp Suite, SQLMap

---

API Security

REST and GraphQL API security testing based on OWASP API Security Top 10.

Capabilities:

• Authentication/authorization testing (BOLA, BFLA, broken auth)
• Injection testing (SQLi, NoSQLi, command injection)
• Mass assignment and SSRF detection
• Rate limiting and resource exhaustion testing
• GraphQL introspection and query depth attacks
• JWT and OAuth vulnerability testing

Trigger phrases: test API security, scan REST API, GraphQL security assessment, OWASP API Top 10

Tools: Burp Suite, Postman, graphql-cop, jwt_tool, Nuclei

---

Cloud Security

Cloud Security

Multi-cloud security assessment for AWS, GCP, and Azure environments.

Capabilities:

• IAM policy analysis and privilege escalation paths
• Storage security (S3, GCS, Azure Blob misconfigurations)
• Network security (VPC, security groups, firewall rules)
• Secrets management and key rotation assessment
• Compliance checking (CIS benchmarks, SOC2, PCI-DSS)
• Serverless security (Lambda, Cloud Functions)

Trigger phrases: assess AWS security, audit GCP permissions, Azure security review, cloud misconfiguration scan

Tools: ScoutSuite, Prowler, CloudSploit, Steampipe, cloud-nuke

---

IaC Security

Infrastructure as Code security scanning for Terraform, CloudFormation, Kubernetes, and Helm.

Capabilities:

• Terraform security with Checkov, tfsec, Terrascan
• CloudFormation scanning with cfn-lint, cfn-nag
• Kubernetes manifest security with kubesec, kube-linter
• Helm chart security analysis
• ARM template scanning for Azure
• Policy-as-code with OPA/Conftest

Trigger phrases: scan Terraform for security issues, check CloudFormation template, Kubernetes manifest security, IaC security audit

Tools: Checkov, tfsec, Terrascan, KICS, kubesec, OPA/Conftest

---

Container Security

Docker and Kubernetes security assessment including image scanning and cluster hardening.

Capabilities:

• Container image vulnerability scanning with Trivy, Grype
• SBOM generation with Syft
• Kubernetes cluster security with Kubescape, kube-bench
• Runtime threat detection with Falco
• Container escape and privilege escalation testing
• Network policy and RBAC analysis

Trigger phrases: scan container image, Kubernetes security audit, check Docker security, cluster hardening assessment

Tools: Trivy, Grype, Syft, Kubescape, kube-bench, kube-hunter, Falco

---

Network Security

Network Pentest

Internal network and Active Directory penetration testing.

Capabilities:

• Network reconnaissance with Nmap, Masscan
• Active Directory enumeration with BloodHound
• Credential attacks (Kerberoasting, AS-REP roasting, password spraying)
• Lateral movement (Pass-the-Hash, Pass-the-Ticket, DCOM, WMI)
• Privilege escalation (DCSync, delegation abuse, GPO attacks)
• Domain dominance (Golden/Silver tickets, skeleton key)

Trigger phrases: internal network pentest, Active Directory assessment, enumerate domain, lateral movement testing

Tools: Nmap, BloodHound, Impacket, CrackMapExec, Responder, Mimikatz, Rubeus

---

Code Security

SAST Orchestration

Static Application Security Testing orchestration with custom rule development.

Capabilities:

• Multi-tool scanning with Semgrep, CodeQL, Bandit
• Custom rule development (Semgrep YAML, CodeQL queries)
• Finding triage and false positive reduction
• CI/CD integration (GitHub Actions, GitLab CI)
• Language-specific scanning (gosec, Brakeman, ESLint security)
• SARIF output and result aggregation

Trigger phrases: scan code for vulnerabilities, write Semgrep rule, triage SAST findings, static analysis setup

Tools: Semgrep, CodeQL, Bandit, gosec, Brakeman, SpotBugs, ESLint

---

SCA Security

Software Composition Analysis for dependency vulnerabilities and supply chain security.

Capabilities:

• SBOM generation (CycloneDX, SPDX formats)
• Multi-ecosystem scanning (npm, pip, Maven, Cargo, Go, Ruby, PHP)
• Vulnerability database correlation (NVD, OSV, GitHub Advisory)
• License compliance checking
• Dependency graph analysis and transitive vulnerability detection
• Supply chain attack detection

Trigger phrases: scan dependencies, generate SBOM, check for vulnerable packages, license compliance audit

Tools: Syft, Grype, Trivy, OWASP Dependency-Check, npm audit, pip-audit, Snyk

---

AI/ML Security

LLM Security

LLM and AI application security testing including prompt injection and guardrail testing.

Capabilities:

• Prompt injection testing (direct and indirect)
• Jailbreaking and guardrail bypass techniques
• System prompt extraction attacks
• RAG pipeline security assessment
• AI agent excessive agency testing
• OWASP LLM Top 10 coverage

Trigger phrases: test LLM for prompt injection, jailbreak AI system, test AI guardrails, RAG security assessment

Tools: Custom scripts, Garak, AI security testing frameworks

---

Security Process

Threat Modeling

Systematic threat identification using industry frameworks and methodologies.

Capabilities:

• STRIDE threat analysis (per element and per interaction)
• PASTA 7-stage methodology
• Attack Tree construction with AND/OR gates
• DREAD risk scoring
• Data Flow Diagram (DFD) creation
• Threat library and mitigation mapping

Trigger phrases: threat model this system, STRIDE analysis, create attack tree, security architecture review

Tools: Microsoft Threat Modeling Tool, OWASP Threat Dragon, draw.io

---

Directory Structure

ai-security-arsenal/
├── README.md
└── skills/
    ├── android-pentest/      # Mobile - Android security
    ├── ios-pentest/          # Mobile - iOS security
    ├── dast-automation/      # Web - Dynamic testing
    ├── api-security/         # Web - API security
    ├── cloud-security/       # Cloud - Multi-cloud assessment
    ├── iac-security/         # Cloud - Infrastructure as Code
    ├── container-security/   # Cloud - Docker/Kubernetes
    ├── network-pentest/      # Network - Internal/AD testing
    ├── sast-orchestration/   # Code - Static analysis
    ├── sca-security/         # Code - Dependency scanning
    ├── llm-security/         # AI/ML - LLM security
    └── threat-modeling/      # Process - Threat analysis

Each skill directory follows the same internal layout — see Skill Architecture.

Quick Start

Mobile Testing

# Android - Connect rooted device and start Frida
adb devices
adb shell "su -c '/data/local/tmp/frida-server -D &'"
# "pentest com.example.app"

# iOS - Connect jailbroken device
iproxy 2222 22
ssh -p 2222 root@localhost
# "pentest com.example.app"

Web Testing

# DAST with Playwright MCP
# "scan https://target.com for vulnerabilities"

# API Security
# "test API security for https://api.target.com"

Cloud/Infrastructure

# Cloud Security
aws configure  # or gcloud auth login
# "assess AWS security for account 123456789"

# IaC Security
# "scan Terraform in ./infrastructure for security issues"

# Container Security
# "scan container image nginx:latest"

Code Analysis

# SAST
# "scan this codebase for vulnerabilities"

# SCA
# "check dependencies for vulnerabilities"

Requirements

Usage with Coding Agents

These skills are designed to work with AI coding agents that support skill/instruction systems. Invoke via natural language:

You: pentest the banking app com.megabank.mobile
Agent: [Executes android-pentest - spawns with Frida, bypasses SSL, extracts data...]

You: scan our AWS account for security misconfigurations
Agent: [Executes cloud-security - runs ScoutSuite, analyzes IAM, checks S3...]

You: check this codebase for vulnerabilities
Agent: [Executes sast-orchestration - runs Semgrep, Bandit, aggregates findings...]

You: threat model the payment processing system
Agent: [Executes threat-modeling - creates DFD, applies STRIDE, generates report...]

Structured output: Each skill emits findings that validate against its schemas/finding.json — pipe directly into SARIF converters, issue trackers, or custom triage pipelines without reformatting.

Parallel execution: Multi-target requests (multiple domains, multiple images, multiple clouds) are handled by spawning sub-agents per target, as documented in each skill’s Sub-Agent Delegation section.

Compatible with: Claude Code, Codex CLI, OpenCode, Aider, and other agents supporting markdown-based skill definitions.

License

For authorized security testing, penetration testing engagements, CTF competitions, and educational purposes only. Always obtain proper authorization before testing.