mobile-security-skill

Mobile Security Skill

Claude Code agents for Android & iOS mobile application security audits, OWASP MASVS/MASTG-aligned, with adversarial exploitation confirmation.

A complete mobile application security assessment toolkit for Claude Code: one orchestrating Skill and 18 specialist subagents, grounded in an OWASP MASTG-style reference checklist, that ends in an adversarial exploitation panel. A prosecutor agent and independent skeptic agents have to actually try to break each other's case before anything is allowed to be reported as Confirmed.

It covers Android and iOS manifest/plist/entitlements, local storage and cryptography, taint analysis (source → sink → sanitizer), WebView/OAuth/OIDC/deep-links/clipboard, Android IPC and UI-layer attacks, iOS platform attacks, anti-tamper/resilience (root/jailbreak/pinning/debugger/Frida detection), network and backend API weaknesses, supply-chain and cross-platform-framework issues (Flutter, React Native, Xamarin, Unity, Cordova, KMP), and business-logic abuse.

This is a dual-use security tool. Read Authorization & Legal before running it against anything.

Jump to: Installation · Why the adversarial panel · Agent roster · Usage · FAQ · Authorization & Legal · Connect

Installation

Option A: Claude Code plugin (recommended)

git clone https://github.com/harshdhamaniya/mobile-security-skill.git

Then in Claude Code:

/plugin install <path-to-cloned-repo>

This picks up .claude-plugin/plugin.json and makes the mobile-security-skill skill and all 18 agents available automatically. Check your Claude Code version's plugin documentation if /plugin install isn't available. The manual method below always works as a fallback.

Option B: Manual copy (works on any Claude Code version)

# Clone wherever's convenient
git clone https://github.com/harshdhamaniya/mobile-security-skill.git
cd mobile-security-skill

# Copy the skill
mkdir -p ~/.claude/skills
cp -r skills/mobile-security-skill ~/.claude/skills/

# Copy the agents
mkdir -p ~/.claude/agents
cp agents/*.md ~/.claude/agents/

Restart Claude Code (or start a new session) so it picks up the new skill and agents. To scope this to one project instead of globally, copy into .claude/skills/ and .claude/agents/ inside that project's root instead of ~/.claude/.

Dependencies

The agents degrade gracefully, but you'll get far more out of them with these on PATH:

• Android: apktool, jadx, aapt2, apksigner, adb
• iOS: plutil, codesign, otool, Xcode command-line tools, xcrun simctl
• Both: frida/frida-tools, objection, a proxy (Burp Suite or mitmproxy)
• Optional but valuable: CodeQL, Semgrep, osv-scanner/dependency-check, trufflehog/gitleaks

None of these are required to start. mobile-recon-triage and every specialist will tell you plainly what it couldn't check because a tool was missing, rather than silently skipping it.

Found this useful? Star the repo. Claude Code/Codex skill directories and plugin marketplaces commonly rank and surface by star count. A star is what makes this findable for the next person doing a mobile security audit. No signup, no cost.

Why the adversarial panel

Static analysis, human or AI, over-reports. A grep hit for addJavascriptInterface or a missing FLAG_IMMUTABLE is a candidate, not a confirmed vulnerability, until someone checks whether a sanitizer elsewhere on the path already closes it, whether the attacker model is actually what it's claimed to be, and whether a platform default on the SDK version in scope already mitigates it. Most "AI finds bugs" tooling stops at the candidate stage and calls it a report.

This skill doesn't. Every finding that matters goes through exploit-verifier-prosecutor, which builds the strongest honest exploitability case it can. Then 2–3 independent exploitability-skeptic instances, each told explicitly to go re-read the code and genuinely try to refute the case rather than rubber-stamp it, get a turn at it. A finding only reaches Confirmed if the skeptics' refutation attempts fail on their own merits. See references/severity-rating.md for the exact decision rule, examples/sample-walkthrough.md for a narrated worked example, and examples/steps-to-reproduce.md for a real run: actual commands, actual agent output, and the actual committed evidence files, not a narrative.

Agent roster