Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
All full reports
Full reportDelivery infrastructureMonograph 34 / 5517 min read

Cloned GeoIP Proxies and APK Workbenches: Modelling Mobile Leads

Ten Nginx edge nodes cloned from one image, a four-architecture staged Android workbench, and seven feed-labeled mobile endpoints

By
Odin by Brandefense
Published
19 November 2026
Reading time
17 min read
Edition
Version 1.1 · TLP:CLEAR
Scope
Prepared November 2026
Cloned GeoIP Proxies and APK Workbenches: Modelling Mobile Leads

1. Executive Summary

In short

Across evidence dated 3 February and 16–20 November 2026, Odin by Brandefense examined ten Nginx reverse proxies in Hong Kong cloned from one server image. It also examined an exposed US workbench holding staged Apktool files for four Android CPU types, and seven feed-labeled mobile endpoints. No shared address, certificate or hash links the three groups, and all 18 host addresses are withheld pending corroboration. Turn off unmanaged APK installs on Android fleets, alert on external APK downloads, and require multi-signal corroboration before blocking cloud hosts.

Mobile trojan campaigns rarely use one server for everything. Regional traffic filtering, app repackaging and command-and-control (C2) often sit with different cloud providers, so a takedown at one layer leaves the others standing. This report looks at one example of each layer, examined from evidence dated 3 February and 16–20 November 2026, and is careful about what links them.

1

Ten Hong Kong reverse proxies share byte-identical files from one server image.

Ten open web directories on Microsoft cloud space in Hong Kong exposed 5,035 files. Every node holds the same custom 14.38 MB Nginx binary, the same GeoIP module and the same 15.44 MB error log. They also share paired certificate bundles; one node also held a mobile download configuration file (path withheld).

2

An exposed workbench holds staged IPTV app files for four Android CPU types.

A US open directory held 5,399 unpacked Apktool files consistent with a streaming lure. Two of the four build trees stage a 12 MiB native library for 32-bit and 64-bit ARM and x86, next to a commercial packer and inline hooking libraries. A staged 65.15 MB APK and two control-panel archives sit at the web root.

3

Seven feed-labeled mobile endpoints remain withheld pending independent behavioral corroboration.

The source totals 138 global sightings across four feed labels: PixPirate, LightSpy, GobRAT and KTLVBackdoor. Seven feed-labeled endpoints appear in the sampled inventory. Two Huawei Cloud endpoints carry a PixPirate feed label and share a certificate name; their C2 role needs independent corroboration. All seven endpoint addresses are withheld.

Three Decisions for Leaders

  1. Stop browser and unmanaged APK installs on Android fleets. Turn off installs from unknown sources on managed Android devices through your MDM. Keep Google Play Protect active and restrict Accessibility Service grants.
  2. Watch for external APK downloads and packed multi-ABI artifacts. Alert on APK files fetched from outside your app store or MDM, and hunt the multi-ABI packer and hooking combinations in Chapter 8.
  3. Keep uncorroborated cloud addresses off block lists. Cloud addresses change owners quickly. Pair the published JARM fingerprints and certificate leads with an independent certificate or endpoint process signal before blocking any host.

The Study in Numbers

NumberWhat it is
10Cloned Nginx reverse-proxy nodes on Microsoft cloud in Hong Kong (5,035 exposed files)
14.38 MBCustom Nginx binary, byte-identical on all 10 nodes, with an 89,976-byte GeoIP module
5,399Unpacked Apktool files on a US Android repackaging workbench
12 MiBSize of a staged native library (name withheld), the same for all four Android CPU types
65.15 MBStaged multi-ABI IPTV APK on the workbench
2Huawei Cloud endpoints with a PixPirate feed label and shared certificate name (addresses withheld)
5Other suspected or candidate feed-labeled mobile endpoints (addresses withheld)

How to use this report.

Security leaders can read Chapters 1, 6 and 7. Mobile, network and cloud engineers should read Chapters 3 to 5. SOC analysts and detection engineers should start with Chapters 8 and 9.

In short

The study covers 18 hosts and 10,434 exposed files in three groups. All 18 host addresses are withheld pending ownership or behavioral corroboration; use structural and TLS hunts instead.

The three groups are a cloned edge fleet, an APK workbench and seven feed-labeled mobile endpoints, which defenders can evaluate using cyber threat intelligence. Together they look like the layers of a mobile trojan delivery chain. But no shared IP address, TLS certificate or file hash links them. Figure 1 therefore shows them as an analytical model, not as one proven operation.

Figure 1: Cloned GeoIP proxies alongside staged APK trees and feed-labeled leads
Figure 1: Cloned GeoIP proxies alongside staged APK trees and feed-labeled leads
GroupSampled hostsPublished addressesExposed files or global feed sightingsHostingStatus
Cloned GeoIP Nginx edge fleet1005,035 exposed filesMicrosoft Corp, Hong KongOwner not established, withheld
Multi-ABI Android APK workbench105,399 exposed filesUnattributed US hostingOwner not established, address and hashes withheld
PixPirate feed-labeled endpoints2018 global PixPirate sightingsHuawei Cloud, ChinaNeeds independent corroboration, addresses withheld
Other feed-labeled mobile endpoints50120 global sightings (GobRAT, LightSpy, KTLVBackdoor)AWS Germany, Alibaba, Huawei, UCloudSuspected or candidate, addresses withheld

Open directories like these can belong to badly configured staging servers or to cloud machines that change tenants. Publishing their addresses, distinctive paths or file hashes could identify an unconfirmed third party, so this report withholds them and focuses on structural detection rules.

3. The Cloned Ten-Node GeoIP Nginx Fleet

In short

All ten Hong Kong nodes share the same binary, module, error-log size and paired certificate bundles. That level of sameness points to one disk image copied ten times; disable directory listing and clear logs before cloning server images.

Many open directories run ordinary distribution Nginx packages and look alike for that reason. These ten are different. They share non-standard file sizes and identical SHA-256 digests across executables, logs, TLS bundles and configuration includes. In total they share 14 file hashes.

The nodes exposed between 99 and 694 files each, 5,035 in all; defenders can use external attack surface management to find similar open directories. The mean is 503.5 files and the median 553.5.

Figure 3: Illustrative GeoIP, APK and C2 stages; no observed links
Figure 3: Illustrative GeoIP, APK and C2 stages; no observed links

Figure 3 is a model. It shows how GeoIP checks, SQL-injection filtering, paired certificates and a download configuration could work together in a mobile delivery chain. We did not observe a victim being redirected or an APK being installed.

One image, three tell-tale files

File (path withheld)NodesSize (bytes)What it is
Custom Nginx binary (path withheld)10 of 1014,379,936Custom Nginx binary, about ten times larger than a stripped distribution build
Dynamic GeoIP module (path withheld)10 of 1089,976Dynamic GeoIP module (ngx_http_geoip_module) that reads country and network variables such as $geoip_country_code and $geoip_org
Pre-filled Nginx error log (path withheld)10 of 1015,440,912Pre-filled Nginx error log, identical on the nine nodes where it was hashed

The error log is the strongest clue. Live proxies write different log lines from the first request. The same 15.44 MB log, hash-identical on the nine nodes where it was hashed, indicates that the machines came from one captured disk snapshot that already held old logs. For defenders, the open directory puts that history on the owner's attack surface.

Paired certificates and configuration files

Every node holds 11 certificate bundles in an SSL folder (path and filenames withheld). Five are portal certificates. The other six form three paired bundles: a base bundle (7,120 to 8,847 bytes) and a companion bundle (7,137 to 9,471 bytes). The real certificate domains and filenames are withheld.

The paired certificate filenames could support domain rotation; no certificate switch or domain suspension was observed. This is a common pattern for grey-market app portals and for malware delivery alike.

Shared file (name and path withheld)NodesSize (bytes)Role
Upstream proxy configuration (path withheld)1 of 1019,834File name suggests upstream proxy mapping; contents and behavior were not verified
Mobile download configuration (path withheld)1 of 1019,123File name suggests app-download routing; contents and behavior were not verified
Companion TLS bundle B (name and path withheld)10 of 109,471Companion TLS bundle; rotation not verified
Companion TLS bundle A (name and path withheld)10 of 107,142Companion TLS bundle; rotation not verified
SQL-injection filter include (path withheld)7 of 106,701Query-string SQL-injection filter include; effect on scanners was not tested
Generated traffic-control config (path withheld)4 of 105,730Upstream routing settings; behavior not verified
Web routing script template (path withheld)1 of 105,487Edge routing template; behavior not verified

The two 19 KB .conf files surfaced on one node, where they were among the 15 largest files. On the other nodes, larger certificate chains pushed smaller .conf files out of the largest-file listing. So "1 of 10" means seen on one node, not missing from nine.

Together, GeoIP checks, an anti-SQL-injection include and a mobile download configuration suggest an edge layer built for traffic filtering and routing. Its effect on automated scanners was not tested. Whether this configuration steered mobile browsers toward a download was not tested.

4. The Multi-ABI Android APK Repackaging Workbench

In short

A US open directory exposes staged Apktool files consistent with IPTV app repackaging across four Android CPU types, alongside a commercial packer and inline hooking libraries. Block unmanaged APK installs on Android fleets.

Sideloaded Android trojans often pose as free streaming or IPTV players, an app impersonation problem that brand protection teams track. Users then install them outside Google Play and grant extra permissions. This workbench shows the server side of that kind of lure: 5,399 files in total. Our review covers the 90 largest files.

Figure 2: Byte-identical edge modules and native packers map to layered rules
Figure 2: Byte-identical edge modules and native packers map to layered rules

Four Apktool trees

The workbench holds four parallel Apktool build trees; their paths are withheld because the host owner is unconfirmed. Four Apktool trees contain files consistent with app repackaging, including branding assets, native libraries and a staged APK; no build was observed.

Three file types point to this unpacking and editing layout. There is an original RSA signing certificate (1,050 bytes), a saved backup of the Android manifest (20,480 bytes), and Apktool duplicate-resource stubs (names and paths withheld).

Staged archives and installer at the web root

Artifact (path withheld)Size (bytes)What it is
Packaged IPTV archive A (.zip)76,546,048Packaged IPTV distribution bundle
Packaged IPTV archive B (.zip)76,546,048Second bundle of the same size
Signed multi-ABI installer (.apk)65,149,417Staged multi-ABI Android installer (65.15 MB)
Control-panel deployment archive (.zip)2,228,822Back-end web control panel archive
Panel rebranding patch archive (.zip)819,308Panel rebranding patch archive

Four CPU types, one packer, two hooking engines

Of the 90 largest files, 63 (70.0%) are native .so libraries. The staged folders cover all four Android NDK ABIs: arm64-v8a, armeabi-v7a, x86 and x86_64. That spans phones, TV boxes and emulators.

  • A fixed 12 MiB native container. In two of the four build trees, one native library (name and path withheld) is exactly 12,582,912 bytes for all four ABIs. The same size on ARM and x86 suggests fixed padding or an appended encrypted container.
  • A DexProtector loader. Each ABI folder stages a DexProtector native loader alongside an encrypted secondary DEX asset (1,339,955 bytes) and an encrypted resource asset (68,720 bytes) in all four trees (names and paths withheld). DexProtector is a commercial app-protection packer.
  • Inline hooking. Next to FFmpeg streaming libraries sit four native hooking libraries, including an older backup copy and a 79,872-byte hook library (names and paths withheld). These libraries are designed to intercept native functions at runtime inside the Android app sandbox.
Workbench file (tree paths withheld)CopiesSize (bytes)Role
Multi-ABI native container library (name and path withheld)812,582,912Native core container, same size for all four ABIs
Primary Dalvik bytecode file (name and path withheld)26,252,104Main compiled Dalvik bytecode
Encrypted secondary DEX asset (name and path withheld)41,339,955Encrypted secondary DEX for DexProtector
Older x86 native hook library (name and path withheld)21,121,444Older x86 native hook binary
Packer unpacking stub, 64-bit ARM (name and path withheld)1349,632Packer unpacking stub, 64-bit ARM
Packer unpacking stub, 32-bit ARM (name and path withheld)1312,296Packer unpacking stub, 32-bit ARM
Native app library (name and path withheld)1282,12064-bit ARM library; behavior unverified
Native inline hook library, 64-bit ARM (name and path withheld)1228,176Native inline hook engine, 64-bit ARM

The file names, file hashes and build-tree paths for these artifacts are withheld because the host owner is unconfirmed. We read the directory listing and file metadata. We did not install or run the APK.

5. Feed-Labeled Mobile Leads: Seven Sampled Endpoints

In short

Seven sampled feed-labeled mobile endpoints appear alongside 138 global feed sightings across four family labels. All seven endpoint addresses are withheld pending independent behavioral corroboration; use the JARM and certificate leads to hunt.

The ten clones show clear internal links. Nothing of that kind ties them to the workbench or to the feed-labeled endpoints below. Figure 4 lays the three groups side by side as a model of how such layers fit together.

Figure 4: Golden-image cloning and Apktool trees expose deterministic hashes
Figure 4: Golden-image cloning and Apktool trees expose deterministic hashes

The source totals 138 global sightings across four feed labels; seven endpoints appear in the sampled inventory. The family names are the feeds' own labels, not our attribution to a named actor.

Two Huawei Cloud endpoints with a PixPirate feed label

Two Huawei Cloud endpoints carry a PixPirate feed label and share a certificate name; their C2 role needs independent corroboration. PixPirate is an Android banking trojan. Three traits appear on the pair.

  1. Same provider. Both sit in a Huawei Cloud data center in China; the two addresses are withheld pending a second qualifying signal.
  2. Same certificate name. The two listeners present the same TLS certificate common name. The name itself is withheld.
  3. Two TLS fingerprints. Each listener has its own JARM fingerprint, shown in the table below. Pair the JARM value with certificate or mobile process signals before blocking.
EndpointHostingJARM fingerprint (hunt only)
Huawei Cloud PixPirate-labeled endpoint A (withheld)Huawei Cloud, China29d29d00029d29d22c29d29d29d29dc2ddcfd203d071c45b4b0ffe3d7b4b89
Huawei Cloud PixPirate-labeled endpoint B (withheld)Huawei Cloud, China21d19d00021d21d21c21d19d21d21d3b0d229d76f2fd7cb8e23bb87da38a20

Five other withheld endpoints

The other three feed labels total 120 global sightings: 55 under GobRAT, 40 under LightSpy and 25 under KTLVBackdoor, alongside 18 under PixPirate. Five endpoints under those three labels appear in the sampled inventory. Each is suspected or candidate only, so its address is withheld.

Feed labelEndpointsHostingStatusShared trait
LightSpy2AWS, GermanySuspectedSelf-signed certificate CN=ADMIN1 and one shared JARM (withheld)
LightSpy1UCloud, Hong KongCandidateSingle signal only
GobRAT1Alibaba Cloud, ChinaSuspectedSingle host on shared cloud space
KTLVBackdoor1Huawei Cloud, ChinaSuspectedSingle host on shared cloud space

CN=ADMIN1 is a default-looking certificate name and turns up on harmless servers too. Treat it as a hunting lead only.

6. Other Explanations and False-Positive Traps

In short

10,434 exposed files and 138 global feed sightings yield hunting leads rather than immediate IP block lists. Pair TLS and file leads with endpoint context before blocking.

Large open directories and single-signal TLS matches are easy to over-read. Figure 5 compares the size of each surface with what this report publishes from it.

Figure 5: Exposed file counts and global feed sightings versus withheld hosts
Figure 5: Exposed file counts and global feed sightings versus withheld hosts

Four other readings of the evidence remain open.

  1. A grey-market app-portal CDN. Cloned proxies with GeoIP modules, anti-SQL includes, paired certificates and mobile download configurations also front offshore gaming and streaming portals. These portals hand out apps outside the stores. No APK behind the download configuration was retrieved, so the ten Hong Kong addresses and their configuration paths stay withheld.
  2. A commercial IPTV white-label builder. The four build trees and panel archives could belong to an illicit IPTV rebranding service that uses DexProtector to protect pirated streams. Whatever the motive, unmanaged APKs with multi-ABI inline hooks bypass Google Play protection. The workbench artifacts need behavioral confirmation before any hash-based block.
  3. Cloud address churn. All 18 hosts sit with large cloud providers: Microsoft, AWS, Huawei, Alibaba and UCloud. Addresses there move between customers quickly. Require an independent behavioral or certificate signal before blocking any cloud address, and review temporary blocks after 30 days.
  4. Default certificates and stock TLS stacks. CN=ADMIN1 and common JARM values recur on benign servers. Suricata SID 9534008 treats CN=ADMIN1 as a hunting signal only.
SurfaceEasy mistakeSafer approach
10 cloned Hong Kong Nginx nodesBlocking cloud addresses or any ngx_http_geoip_module buildKeep the addresses off block lists; hunt the GeoIP module paired with multi-PEM TLS configs
APK workbenchFlagging legitimate CI apktool or FFmpeg buildsRequire multi-ABI native libraries, packer assets and hooking libraries together
CN=ADMIN1 certificate and JARMMatching default appliance certificates or stock TLS stacksHunt only, and confirm with mobile endpoint process data
PixPirate-labeled Huawei Cloud pairBlocking on a feed label and JARM without a second signalWithhold from block lists; pair JARM with certificate or mobile process telemetry

7. What Defenders Should Do

Mobile and endpoint

  • Turn on the Android Enterprise policies that block installs from unknown sources through your MDM. Keep Google Play Protect on, and limit Accessibility Service rights for apps from outside the store.
  • Test the three unvalidated YARA structural hunts from Chapter 8 on configuration text, certificate manifests and multi-ABI build specifications before adding them to sandbox pipelines.

Network security

  • Use the two PixPirate-labeled JARM fingerprints and the CN=ADMIN1 certificate subject as TLS hunting leads on firewalls and mobile ZTNA gateways. Require certificate or endpoint process corroboration before blocking.
  • Alert on APK files (application/vnd.android.package-archive) delivered from outside approved stores and MDM hosts.

Cloud and platform

  • Check your own Nginx reverse proxies. Turn directory listing off (autoindex off) and block public access to binary, module, certificate, log and configuration directories, as well as *.conf and *.inc files.
  • Clear old logs from golden VM images before you clone them.

Detection engineering

  • Deploy the Suricata, Sigma and KQL rules in Chapter 8. Fill the MDM and Android CI allowlists with exact hostnames first, so authorized builds do not flood the queue.

8. Detection Rules

The full rules are in the Detection rules tab. The Suricata file passes a syntax lint. The YARA, Sigma and KQL files were checked by reading, not loaded in live engines for this edition. None of the rules was replayed against live traffic, so test them in your own environment first.

Rule fileWhat it catches
odin_cloned_geoip_nginx_apk_delivery_and_pixpirate.rulesSuricata 7: five signatures for exposed SQL-filter regexes, GeoIP module downloads, external APK deliveries, unpacked multi-ABI .so requests and CN=ADMIN1 certificates
sentinel_geoip_apk_delivery_and_mobile_c2_hunt.kqlMicrosoft Sentinel and Defender XDR: explicit .apk URL downloads outside approved MDM hosts, plus nginx, apktool, java or adb commands that name the GeoIP module or multi-ABI repackaging markers
sigma_exposed_nginx_conf_geoip_and_apk_sideload.ymlSigma (3 rules): proxy requests for external APKs or unpacked multi-ABI .so paths; unmanaged nginx launches with the GeoIP module; and Linux repackaging tool command lines referencing Android NDK ABI names
odin_cloned_nginx_apk_conf_and_jni_library.yarYARA (3 unvalidated structural hunts): Nginx configuration text, certificate manifests and multi-ABI build specifications.

How the Suricata signatures work

SIDLooks forWhy
9534004An HTTP 200 response body holding select|union and $query_stringCatches an exposed SQL-injection filter include file
9534005A GET for ngx_http_geoip_module.soCatches retrieval of the GeoIP module from outside
9534006An external response with content type application/vnd.android.package-archive and a ZIP headerHunt only: flags APK deliveries outside approved channels
9534007A GET for any .so library under /lib/<abi>/ across the four Android NDK ABIsHunt only: flags retrieval of unpacked multi-ABI Android native libraries
9534008A TLS server certificate with CN=ADMIN1Hunt only: default-looking certificate seen on hosts with a LightSpy feed label

Keep SIDs 9534004 through 9534008 in alert mode, with pass rules for approved MDM and build hosts.

How the other rules work

The KQL query looks back 7 days. Because the unconfirmed hosts' paths and uncorroborated C2 addresses are withheld, the network branch matches the generic .apk URL pattern explicitly ((?i).apk(?:?|$)). The query skips remote hosts in an exact MDM hostname list and devices in an exact build-host list.

The process side looks for nginx, apktool, java or adb with the GeoIP module or Android NDK ABI and repackaging markers on the command line. Each match gets one hunting label: hunt a multi-ABI APK workbench, hunt a cloned GeoIP edge config, or hunt an unmanaged APK transfer.

The first Sigma rule reads web proxy logs and flags requests for .apk URLs or unpacked .so libraries under /lib/<abi>/. Three exact MDM hostnames and one internal scanner address are excluded. The second flags nginx launches whose command line names ngx_http_geoip_module.so. The third pairs an unvalidated selection_repack_tool executable list with the Android NDK ABI names arm64-v8a, armeabi-v7a or x86_64. The third rule flags Linux command lines mentioning these ABI names under repackaging tools; analysts must verify that a repackaging process ran.

YARA ruleFires when
ODIN_Cloned_Nginx_GeoIP_APK_Delivery_Configngx_http_geoip_module.so and proxy_pass appear with $geoip_country_code, $geoip_org or the select|union and $query_string filter terms
ODIN_Cloned_Nginx_Multi_PEM_TLS_ManifestMultiple ssl_certificate and ssl_certificate_key directives appear with .pem
ODIN_MultiABI_APK_Repack_Build_Specapktool or AndroidManifest.xml appears with all four Android NDK ABI names (arm64-v8a, armeabi-v7a, x86 and x86_64)

The two MITRE ATT&CK for Mobile tactics in play are Initial Access (TA0027), through sideloaded and repackaged apps, and Command and Control (TA0037), through encrypted channels.

9. Indicators

The Indicators tab holds 0 published block indicators: all 18 host addresses and all 18 file hashes are withheld pending independent behavioral corroboration or host ownership checks. Three hunting indicators appear in this report only: the two PixPirate-labeled JARM fingerprints and the CN=ADMIN1 certificate subject.

TypeCountWhere to use it
Network addresses0 (18 withheld)Withheld pending independent behavioral corroboration and host ownership checks
File hashes (SHA-256)0 (18 withheld)Withheld pending host ownership and behavioral confirmation
JARM fingerprints2 (1 withheld)TLS hunting only, pair with certificate or mobile process signals (Chapter 5)
Nginx configuration paths0 (withheld)Owner-specific configuration paths withheld pending host ownership confirmation
Certificate subject1TLS hunting only: CN=ADMIN1

Withheld: all 18 host addresses, covering the ten Hong Kong nodes and the US workbench. This also includes the two Huawei Cloud PixPirate-labeled endpoints and the five other suspected or candidate feed-labeled hosts. Also withheld are all 18 file hashes from the unconfirmed open-directory hosts (eight from the Hong Kong fleet and ten from the US workbench). The owner-specific Nginx configuration paths, the distinctive workbench paths and file names, the real certificate domains and filenames, the shared PixPirate certificate name and the LightSpy JARM are withheld as well.

10. Scope and Limits

  • Evidence dates. The findings draw on evidence dated 3 February and 16–20 November 2026, with a research cutoff of 20 November 2026.
  • Files, not outcomes. We read directory listings and file metadata. We did not run any binary or install the APK, and we saw no victim device.
  • No proven chain. No shared address, certificate or hash links the Hong Kong fleet, the US workbench and the feed-labeled endpoints. The end-to-end delivery picture is an analytical model.
  • Owners not established. The Hong Kong fleet could serve a grey-market app portal, and the workbench could serve an IPTV piracy business. Their addresses, configuration paths and host-derived file names are withheld for that reason.
  • Feed labels are not attribution. PixPirate, LightSpy, GobRAT and KTLVBackdoor are the feeds' own family labels.
  • Partial view. The workbench figures describe its 90 largest files out of 5,399. The Hong Kong figures rely on each node's largest files, so a file missing from a node's list may still exist there.
  • Rule testing. The Suricata file passes a syntax lint. The other rule files were not loaded in live engines for this edition, and no rule was replayed against live traffic. Test and tune them before production use.

11. Sources

  1. NGINX documentation. Module ngx_http_geoip_module. https://nginx.org/en/docs/http/ngx_http_geoip_module.html (verified 20 November 2026)
  2. NGINX documentation. Module ngx_http_proxy_module. https://nginx.org/en/docs/http/ngx_http_proxy_module.html (verified 20 November 2026)
  3. Android Developers. Android ABIs. https://developer.android.com/ndk/guides/abis (verified 20 November 2026)
  4. Android Open Source Project. Application Sandbox. https://source.android.com/docs/security/app-sandbox (verified 20 November 2026)
  5. OWASP Mobile Application Security. Mobile Application Security Testing Guide (MASTG). https://mas.owasp.org/MASTG/ (verified 20 November 2026)
  6. MITRE ATT&CK for Mobile. TA0027 Initial Access. https://attack.mitre.org/tactics/TA0027/ (verified 20 November 2026)
  7. MITRE ATT&CK for Mobile. TA0037 Command and Control. https://attack.mitre.org/tactics/TA0037/ (verified 20 November 2026)
  8. Suricata documentation. HTTP Keywords. https://docs.suricata.io/en/latest/rules/http-keywords.html (verified 20 November 2026)

What is…?

Short explanations of the tools and families named in this study.

All explanations

Frequently asked questions

What did we find?

The study covers 18 hosts and 10,434 exposed files in three groups. All 18 host addresses are withheld pending ownership or behavioral corroboration. A live proxy writes its own log from the first request. The log was hash-identical on nine nodes where hashes were available, pointing to one disk image copied across the fleet.

How do the cloned proxies work?

The nodes pair Nginx's GeoIP module with an anti-SQL-injection filter include (seen on seven of ten) and, on one node, a mobile download configuration file (path withheld). The GeoIP module lets the server decide what to show based on the visitor's country and network. The include appears designed to filter SQL-injection requests; we did not test whether it blocked scanners. Whether this configuration steered mobile browsers toward a download was not tested.

What does the workbench hold?

The workbench holds staged files for one streaming lure. Four Apktool trees contain files consistent with app repackaging, including branding assets, native libraries and a staged APK; no build was observed. Their build-tree paths are withheld because the host owner is unconfirmed.

Why are all 18 host addresses withheld?

Large cloud providers move addresses between customers quickly. A block on an uncorroborated address can hit an innocent tenant a week later. Two Huawei Cloud endpoints carry a PixPirate feed label and share a certificate name, but their C2 role needs independent corroboration. We therefore withhold both addresses and publish their JARM fingerprints as hunting leads only.

Defender kit

Indicators of compromise

Verified adversary infrastructure and tool hashes from this study. Every value below is also discussed in the report.

No network or file indicators are published for this study

The hosts in this report are exposed or unattributed systems rather than confirmed adversary infrastructure, so they stay anonymous. Use the detection rules below to find the same exposure in your own estate.

28 records from this study are not published in this table: candidate, suspected and dual-use hosts, victim-related values, and hashes of exposed secrets or files.

Defender kit

Detection rules

The rule files exactly as printed in the report. Test them in your own engine before deployment.

Suricata odin_cloned_geoip_nginx_apk_delivery_and_pixpirate.rules 15 lines
Download
YARA odin_cloned_nginx_apk_conf_and_jni_library.yar 62 lines
Download
Sentinel KQL sentinel_geoip_apk_delivery_and_mobile_c2_hunt.kql 39 lines
Download
Sigma sigma_exposed_nginx_conf_geoip_and_apk_sideload.yml 96 lines
Download