Skip to main content
Best-effort symbol/string survey of the native libraries. The first pass surveyed the 7 most relevant libraries, plus one targeted ARM64 disassembly of libxlua.so’s CalcChatSign helper (see the feasibility table below and Findings & further work). The 1.0.364 coverage pass surveyed all 39 libraries across both ABIs by symbols, strings, and P/Invoke maps, with two more targeted arm64 disassemblies. libsmsdk.so (ShuMei) is the one library that has since been fully decompiled — its Obfuscator-LLVM string encryption was reversed by emulating the .datadiv_decode* init functions, and the resulting device-profile/deviceprofile protocol is documented in ShuMei device fingerprint. The question that matters: does anything below the protocol layer block a clean-room Go client? For the one existing account tested against production, the answer is no, see the Warning below for the full scope.

libil2cpp.so is a Mono bridge, not an IL2CPP build

Only 1.8 MB, small for a Unity IL2CPP binary, and for good reason: it contains zero strings for AESHelper, NetworkManager, SendLuaMessage, or even generic Unity type names. Assembly-CSharp is never AOT-compiled here at all; everything gameplay-relevant lives in the hot-update .mdl tier. What runs that tier is FunPlus’s FM_Mono bridge, not HybridCLR as earlier versions of these docs said (static-analysis-only):
  • In both ABIs, libil2cpp.so exports the il2cpp_* ABI plus FM_Mono::*, LoadAssemblyByName, AddAssemblySearchPath(…, NativePersistType), SetAssemblyOnlyReadFromPackage, LoadAssemblyWithImageBinary, and visitNoUpdateAssemblies.
  • It implements them on upstream Mono 6.13.0 (libmonoboehm-2.0, the same build in both ABIs). There’s no hybridclr or huatuo symbol or string in any of the 39 libraries.
  • The Java launcher com.rivergame.fm_mono.FMMonoBaseActivity loads it, and the managed side is FM_MonoLib.dll.
  • It adds persistentDataPath/Assemblies/ as a search path for hot-updated .mdl files and skips the persistent copy for any name in noUpdateAssemblies.json, which lists only SmartFox2X.
The .mdl XOR scheme itself is unchanged, see Extraction methodology.

Feasibility table

MechanismWhere it enters the wire protocolBlocking risk
ShuMei device fingerprint (boxId)login field shumeiBoxIdNone identified, live-tested (see Live validation against production): empty/synthetic values are accepted without issue for a brand-new guest account; reconnecting into an established account initially looked ShuMei-gated but turned out to be an unrelated packageName/platform/appVersion identity mismatch against the access token. Once that was fixed, reconnect worked, and it kept working against the shipped code after a later security fix replaced part of the request with placeholders (August 2026; since the 1.0.370 follow-up the iOS ta fills lw_device_id/lw_airKey from the Login and lw_shumei_id from the cached shumeiId, or replays a captured ta; a live config-copy test is pending before the cron uses it: a copy without a shumeiId takes two runs, the first saving the pushed id and the second sending it, while a copy that already has one, such as a recapture with pcap -session-out, sends it from its first run). On 2026-10-04 a captured iOS session also reconnected with shumeiBoxId sent empty, see Live validation against production. No ShuMei-specific block was ever found. Fully decompiled (Obfuscator-LLVM strings reversed); the device-profile protocol is documented in ShuMei device fingerprint
Tencent ACE anti-cheat reportNot found on the game-server protocol at all, local IPC to Tencent’s own backendNone identified
SecurityCode/OneCode/CoreVlogin request, ~3 fieldsNone: plain MD5 with a hardcoded salt, fully reproducible
packageSign (despite the name, not a cert hash, see Findings & further work)login requestNone: actually SHA1(packageName), a constant public string; trivially reproducible without touching the APK’s real signature. The field that genuinely depends on the signing cert is psh (see Identity, login & session), and the Go client reproduces the Android formula correctly; 1.0.364 is signed with the same certificate, so that value is unchanged. In iOS mode it sends the iOS app’s own formula, recovered from the iOS binary’s rg_getPsh:; see Identity, login & session. Until 2026-10-18 iOS mode sent the Android value, which the server accepted for weeks.
libxlua.so CalcChatSign/AES/MD5 helpersNone in 1.0.364: the native chat branch in ChatNetManager:InitWebSocketInst is unreachable, so chat always uses the C# ChatServiceNone, resolved via direct ARM64 disassembly (see Findings & further work): plain unkeyed MD5(MD5(appId+uid)+timestamp), byte-identical to the known C# formula, no custom key material. Only the armeabi-v7a libxlua registers rg_websocket; arm64’s luaopen_native registers only rg_sqlite, rg_async, cutils, md5, base64, and crypt, and LuaEntry:__InitCModule(), the only code that would expose the native socket, is never called (static-analysis-only)
TLS / certificate validationAll HTTPS trafficNone: no pinning evidence found anywhere. Standard chain validation everywhere except the XAssetPro bundle downloader, which disables certificate validation (see Startup, version-check & hot-update)
Google Play Integrity / SafetyNet, Not used at all in this build
Bottom line, updated after live testing, see Live validation against production. Static analysis alone said nothing here blocks impersonation, and live testing against production confirms it fully for guest accounts: nothing blocks a pure Go client from creating and driving new guest accounts or from completing the email-verification bind of a guest session to a real account, both work end-to-end, repeatedly, against the real account used for testing. Reconnecting into an already-established account’s live game state also worked end-to-end at the time it was tested, but that test used the ta analytics blob’s real captured device/anti-fraud sub-fields; a later security fix (see Live validation against production) replaced those sub-fields with placeholders, and reconnect is now re-confirmed still working with those placeholders, an unattended cron reconnected and collected real resources over multiple days (August 2026), so the server does not require ta’s real device sub-fields. The reconnect step initially looked like a genuine protocol-level session gate (reproducibly rejected, even replaying that account’s own genuine device credentials from a real prior login), but turned out to be a packageName/platform/appVersion/versionCode identity mismatch against the access token, not a security check, see Live validation against production for the full methodology, test matrix, and the current state of the reconnect open question.

ABI parity

The two Android ABIs don’t ship the same native code (1.0.364; static-analysis-only):
  • arm64-v8a ships 21 libraries and armeabi-v7a ships 18. Tencent GME voice (libgmesdk, libtxffmpeg, libtxsoundtouch) is arm64-only, and the C# has no ABI guard around it.
  • libxlua is built from different source trees per ABI: the WebSocket code exists only in v7a, and the marshalling, ZLib, and async exports exist only in arm64.
  • libsqlite3 is 3.28.0 on arm64 and 3.19.3 on v7a, and libxlua embeds its own SQLite 3.38.0.
  • The engine is Unity 2019.4.41f1 (from libunity.so), while a Java crash string says 2019.4.40f1 and the C# hard-codes unityVersion=440 in check-version requests. goclient sends the same 440; keep it rather than “fixing” it to 441.