Firmware-level Android Backdoor via libandroid_runtime Zygote Injection
Overview
Supply-chain tampering of /system/lib[64]/libandroid_runtime.so can hijack android.util.Log.println_native so that every app forked from Zygote executes attacker code. The Keenadu backdoor adds a single call inside println_native that drives a native dropper. Because all app processes run this code, Android sandbox boundaries and per-app permissions are effectively bypassed.[1]
The same post-root execution model also appears in multi-stage Android rootkits delivered from apparently benign Play-distributed apps. In McAfee’s Operation NoVoice analysis, the malware first landed in user space, obtained root with device-tailored exploits, and then replaced libandroid_runtime.so and libmedia_jni.so on the system partition so that every app spawned by zygote inherited the attacker hooks after reboot.[3]
Recent supply-chain patterns (2025+)
Recent Triada-linked investigations show the same “infect Zygote once, reach every app” outcome being delivered through counterfeit/pre-infected firmware and, in some cases, through OTA-related system components instead of a purely post-sale root step.[4] Two practical consequences:
- Compare
ro.build.fingerprintwith vendor-published factory images. In recent counterfeit-device cases, near-match fingerprints were a useful clue that the device was not running a legitimate build. - Hunt beyond
libandroid_runtime.so. Recent firmware families have also used privileged system APKs / services (launcher, facial-recognition utilities, OTA/update paths) as bootstrap or fallback footholds, so a “clean” shared library does not fully rule out a supply-chain compromise.
Pre-root staging: SDK init hijack + PNG tail polyglot
Before the libandroid_runtime.so replacement, NoVoice used an app-level bootstrap that is worth recognizing during APK triage:[3]
- Auto-exec on first launch: bootstrap code ran from a tampered Facebook SDK init path, so no extra user interaction or suspicious permission prompt was required.
- Payload smuggling in assets: the app shipped a valid PNG with an encrypted blob appended after the PNG
IENDmarker. Android image decoders ignore trailing bytes, but the loader carved the tail intoenc.apk, decrypted it intoh.apk, loaded it, and deleted the staging files. - Triage clue: if bytes immediately after
IENDbegin with a magic such asCAFEBABE, assume the image is acting as a polyglot carrier for a Java class / JAR / APK payload rather than as a pure media asset.
Quick checks:
pngcheck -v suspicious.png
tail -c +1 suspicious.png | xxd | tail
binwalk -e suspicious.png
Hunting notes:
- Search APK
assets/for PNGs with unexpected trailing bytes or high entropy afterIEND. - Trace early init paths such as
Application.onCreate, third-party SDK bootstrap code, or nativeJNI_OnLoadhandlers that open asset PNGs and then write*.apk/*.jarfiles into the app sandbox.
Dropper path: native patch → RC4 → DexClassLoader
- Hooked entry: extra call inside
println_nativeto__log_check_tag_count(injected static liblibVndxUtils.a).[1] - Payload storage: RC4-decrypt blob embedded in the
.so, drop to/data/dalvik-cache/arm[64]/system@framework@vndx_10x.jar@classes.jar. - Load & execute:
DexClassLoaderloads the jar and invokescom.ak.test.Main.main. Runtime logs use tagAK_CPP(triage artifact). - Anti-analysis: aborts in Google/Sprint/T-Mobile system apps or if kill-switch files exist.
- Zygote role split:
- In
system_server→ instantiateAKServer. - In any other app → instantiate
AKClient.
- In
What changes on newer Android releases?
- In AOSP, legitimate
android.util.Log.println_nativeis a minimal JNI wrapper whose relevant job is calling__android_log_buf_write. When auditing a vendor build, any extra helper call or control-flow branch around this path deserves immediate review. - On Android 14+, apps using dynamic code loading must mark
DEX/JAR/APKpayloads read-only before loading. Older rootkit chains that drop a writable jar to disk therefore stand out more clearly on modern targets.[5] - A common adaptation is to keep the final payload off disk and switch the last stage to
InMemoryDexClassLoader. If you confirm the zygote hook but never recover a stableclasses.dex, assume the decrypted DEX may only exist in aByteBufferat runtime. For generic fileless DEX hunting, see this Android malware post-exploitation page.
Binder-based client/server backdoor
AKServer(running insystem_server) sends protected broadcasts:[1]com.action.SystemOptimizeService→ binder interface for clients.com.action.SystemProtectService→ binder interface for downloaded modules.
AKClient(inside every app) receives the interface via broadcast and performs anattachtransaction, handing an IPC wrapper so the server can load arbitrary DEX inside the current app process.- Exposed privileged operations (via
SystemProtectService): grant/revoke any permission for any package, retrieve geolocation, and exfiltrate device info. This centralizes privilege bypass while still executing code in chosen target apps (Chrome, YouTube, launcher, shopping apps, etc.).
C2 staging, crypto, and gating
- Host discovery: Base64 → gzip → AES-128-CFB decrypt with key
MD5("ota.host.ba60d29da7fd4794b5c5f732916f7d5c"), IV"0102030405060708".[1] - Victim registration: collect IMEI/MAC/model/OS, encrypt with key
MD5("ota.api.bbf6e0a947a5f41d7f5226affcfd858c"), POST to/ak/api/pts/v4with paramsm=MD5(IMEI)andn=w|m(network type). Responsedatais encrypted identically. - Activation delay: C2 serves modules only after ~2.5 months from an “activation time” in the request, frustrating sandbox detonations.
- Module container (proprietary):
struct KeenaduPayload {
int32_t version;
uint8_t padding[0x100];
uint8_t salt[0x20];
KeenaduChunk config; // size + data
KeenaduChunk payload; // size + data
KeenaduChunk signature;// size + data
} __packed;
- Integrity: MD5 file check + DSA signature (only operator with private key can issue modules).
- Decryption: AES-128-CFB, key
MD5("37d9a33df833c0d6f11f1b8079aaa2dc" + salt), IV"0102030405060708".
Post-root persistence: wrapper libraries + framework patching + self-heal
Once root is available, replacing libandroid_runtime.so is only one layer of persistence. NoVoice shows a more resilient pattern:[3]
- Wrapper replacement instead of inline patching: the installer backed up the original system library and replaced it with an architecture-matched hook wrapper. The same campaign also replaced
libmedia_jni.so, giving multiple code-execution choke points inside the framework. - Second-stage persistence in framework bytecode: after the library swap, a dedicated patcher modified pre-compiled Android framework bytecode on disk. This means restoring the original
.somay still leave injected redirections active. - Self-healing watchdog: a daemon checked the installation roughly every 60 seconds, restored missing components, and could force a reboot if reinsertion kept failing. The malware also replaced the system crash handler / recovery flow so rebooting re-launched the rootkit.
- Per-app payload assembly at runtime: after reboot, the replaced
libandroid_runtime.socaused each spawned app to load attacker code. NoVoice stored secondary payloads as fragments inside the malicious library, assembled them in memory, and deleted the disk copies immediately after load.
Practical implications:
- Factory reset is insufficient when the malware has modified the system partition or framework artifacts. Reflash the firmware instead.
- Diff both
/system/lib*/libandroid_runtime.soand framework oat/odex/vdex artifacts; checking only the shared library can miss the bytecode persistence layer. - If a device keeps restoring the malicious library after manual cleanup, look for a watchdog daemon, modified recovery scripts, or a replaced crash-handler path that is re-seeding the implant on boot.
Persistence & forensic tips
- Supply chain placement: malicious static lib
libVndxUtils.alinked intolibandroid_runtime.soduring build (e.g.,vendor/mediatek/proprietary/external/libutils/arm[64]/libVndxUtils.a).[1] - Firmware auditing: firmware images ship as Android Sparse
super.img; uselpunpack(or similar) to extract partitions and inspectlibandroid_runtime.sofor extra calls inprintln_native.[2] - On-device artifacts: presence of
/data/dalvik-cache/arm*/system@framework@vndx_10x.jar@classes.jar, logcat tagAK_CPP, or protected broadcasts namedcom.action.SystemOptimizeService/com.action.SystemProtectServiceindicate compromise.[1] - For Android rootkits deployed from apps instead of factory supply chain, also inspect:[3]
- APK
assets/for polyglot PNGs with appended data afterIEND - Replaced
/system/lib*/libandroid_runtime.soor/system/lib*/libmedia_jni.so - Modified framework oat/odex/vdex files that preserve hook redirections
- Periodic watchdog processes that rewrite removed files or trigger forced reboots
- APK
- Compare the device
ro.build.fingerprintand hashes of privileged system APKs / libraries against vendor-published firmware. In recent counterfeit-device cases, subtle fingerprint drift was one of the few early hints that the image was not authentic.[4] - Because the legitimate
println_nativeimplementation is tiny, added action strings such ascom.action.SystemOptimizeService,com.action.SystemProtectService, theAK_CPPlog tag, or stray loader paths likesystem@framework@vndx_10x.jar@classes.jarare high-signal triage artifacts.[1]
Quick rooted-device triage
adb shell getprop ro.build.fingerprint
adb shell 'for f in /system/lib*/libandroid_runtime.so /system/lib*/libmedia_jni.so; do [ -e "$f" ] && sha256sum "$f"; done'
adb pull /system/lib64/libandroid_runtime.so . 2>/dev/null || adb pull /system/lib/libandroid_runtime.so .
strings -a libandroid_runtime.so | rg 'AK_CPP|com.action.SystemOptimizeService|com.action.SystemProtectService|vndx_10x|com.ak.test.Main'
adb shell logcat -d | rg 'AK_CPP|SystemOptimizeService|SystemProtectService'
If the firmware ships as super.img, extract the dynamic partitions first (simg2img when needed, then lpunpack) and diff the resulting system*.img contents against a trusted factory image.