DEV Community

정상록
정상록

Posted on

Claude Code Autonomously Hacked Subway Surfers on a Rooted Android Phone

TL;DR

Workers IO connected Claude Code to a rooted Android phone with three tools (UI Automator, mitmproxy, Frida). The AI agent autonomously hacked Subway Surfers — going from 10,000 coins to 2,050,000 in two game sessions. The most interesting part wasn't the hack itself, but the agent's ability to pivot strategy mid-session.

The Setup

Three tools via ADB on a rooted Android:

UI Automator  → Screen reading (XML dumps)
mitmproxy     → Network traffic interception
Frida         → Runtime hooking (JavaScript injection)
Enter fullscreen mode Exit fullscreen mode

The agent ran a continuous loop: observe → act → intercept → decide → repeat.

The Strategy Pivot

Here's what makes this actually interesting.

Claude Code started with network interception — standard first move for mobile pentesting. But Subway Surfers is a Unity game. Gameplay is entirely client-side.

The agent noticed zero game-relevant traffic and made a decision:

"Network proxy is meaningless. Switch to IL2CPP binary analysis."

That's adaptive reasoning. Same thing a human pentester does when their initial approach hits a dead end.

The Technical Path

APK Analysis 

base.apk                    → 209MB
split_config.arm64_v8a.apk  → 32MB (native libs)
global-metadata.dat          → 15MB (IL2CPP metadata)
libil2cpp.so                 → 79MB (compiled game binary)
Enter fullscreen mode Exit fullscreen mode

The global-metadata.dat file is the goldmine for IL2CPP games. It contains original C# class names, method names, and field offsets — all in plaintext. This is a structural weakness of Unity IL2CPP builds.

Key Classes Discovered 

WalletModel           // Currency management
RunSessionData.AddCoins  // Coin collection during runs
SafeInt               // Anti-cheat wrapper
CurrencyType          // Enum: Coins=1, Keys=2, Hoverboards=3
Enter fullscreen mode Exit fullscreen mode

SafeInt Anti-Cheat 

actual_value = _value - _offset  // XOR-based obfuscation
Enter fullscreen mode Exit fullscreen mode

Effective against memory scanners (GameGuardian etc.). Completely useless against Frida's function-level hooks.

The Exploit: Double Amplification 

// Pseudocode of what the agent wrote
Interceptor.attach(AddCoins, {
  onEnter: (args) => { args[1] *= 100; }  // 100x at collection
});

Interceptor.attach(AddCurrency, {
  onEnter: (args) => { args[1] *= 100; }  // 100x at wallet save
});
Enter fullscreen mode Exit fullscreen mode

Result:

Event Coins
Start 10,000
Collect 4 coins (4 × 100) +400
End-of-run bonus +400
Wallet save (400 × 100) +40,000
After 1 run 50,000
After 2 runs 2,050,000

Why This Matters

Subway Surfers is single-player — no real harm. But the vulnerabilities exposed are everywhere:

  1. Client-side trust — No server validation on game state
  2. Metadata exposure — IL2CPP keeps original symbols accessible
  3. No runtime protection — Frida injection goes undetected

These same patterns exist in fintech apps, e-commerce checkouts, and authentication flows.

The Broader Trend

This isn't an isolated case:

  • Zane St. John: Claude Code reverse-engineered Android projector malware
  • Disney Infinity RE: 13-year-old game binary reversed in 24 hours
  • Randy Westergren: Claude + Frida bypassed Flutter SSL pinning
  • RAPTOR: Autonomous security research framework built on Claude Code

Try It 

npx skills add workersio/spec
Enter fullscreen mode Exit fullscreen mode

Requires rooted Android + ADB. Test only on your own devices.

Source: Workers IO Blog

What's your take — does AI-powered pentesting lower the bar for defenders or attackers?

Top comments (0)