All research

AetherC2 - "Yad2 Delivery" Android Attack-Chain Analysis

A four-stage Android banking trojan delivered to Israeli Yad2 sellers as a fake delivery app. We recovered the operator's mTLS certificates from the sample and watched a live botnet of 44 victims read-only: SMS and OTP theft, a live keylogger, covert remote control, and a SOCKS5 proxy that routes bank fraud through the victim's own phone.

  • mobile malware
  • threat intel
  • android
  • banking trojan
Confidential. This documents a live malware operation and its command server. Every sample below is redacted - the JSON structure is real; victim names, phone numbers, IP addresses, keystroke text, operator notes and screen contents are replaced with [REDACTED]. All intelligence was collected passively, read-only. No command was ever sent to any victim device.

The malware reaches Israeli Yad2 sellers as a fake "delivery/courier" app. It is a four-stage packed banking trojan - AetherC2 - operated over MQTT at ssl://194.5.97.27:443 with mutual TLS.

Using certificates and credentials extracted statically from the sample, passive read access to the command server confirmed a live botnet of 44 Israeli victims with SMS and OTP theft, a live keylogger, covert remote control, web-inject, and an exposed operator-panel feed whose notes are written in Russian. Campaign tag: yad2.


How the attack works

The short version, before the technical detail.

  1. The lure. A scammer sends the seller a fake "a buyer was found for your item" message impersonating Yad2, with a link to a counterfeit app-store page.
  2. The fake update. The app looks normal, then shows a fake "Google Play Services update" screen that pressures the victim into granting powerful permissions - Accessibility above all.
  3. The hidden install. Behind the decoy it quietly unpacks and installs the real banking trojan in several hidden layers, so scanners see only a harmless-looking app.
  4. The takeover. The trojan connects to the attacker's server and can now read SMS and one-time codes, record keystrokes, and watch and control the screen invisibly.
  5. The fraud. A Russian-speaking operator opens the banking app on the victim's own phone, uses the stolen SMS codes to pass two-factor authentication, and empties the account.

Initial access - the phishing lure

The seller receives an SMS claiming an item has been bought and that a separate "YAD2 Delivery" app is needed to confirm the order. The link - https://yad2appl.org/app-download - opens a page crafted to look like a Google Play listing: verified badge, 4.8 stars, a million downloads, 128,000 reviews. Tapping Install sideloads the APK.

A counterfeit Google Play listing for "Yad2 Delivery" on attacker infrastructure
The fake Play page at yad2appl.org - attacker infrastructure, not Google's

Infection and unpacking chain

In plain terms: the app the victim installs is a harmless-looking shell wrapped in four layers. Each one silently decrypts and installs the next, until the real banking trojan sits on the device. The layers and the encryption exist for one purpose - to hide it from antivirus and scanners.
StageWhat it isCryptoMechanism
1APK (decoy)—React-Native UI plus the loader Hhopetomato
1 → 2DropperRC4Reads a junk-path asset ending dspePLn.otf, decrypts, unzips, injects the DEX and swaps the live Application through ActivityThread reflection
2 → 3PackerAES-256-GCMStage 3 is appended to assets/banner_48222095.bin behind a fake WEBP header. Trailer is a length at len-6 plus magic 0xAE7D; the key lives in TokenMap as Base64 then XOR. Installed silently with SessionPackageInstaller
3 → 4TrojanRC4Same loader technique from asset DPSlSR.html. The native libraries libyaan7ldva.so and libztx033yop.so ship inside the stage-3 APK

Anti-analysis runs throughout: a malformed ZIP, deliberately wrong CRCs, a fake "encrypted" flag, a STORED manifest whose compressed size is smaller than its uncompressed size, zero padding, and a decoy DEX.


Command-and-control architecture

In plain terms: every infected phone and the attacker's dashboard talk to one central server over an encrypted, chat-like bus. Phones push stolen data up; the server sends commands down.

The broker is Mosquitto 2.0.15 on 194.5.97.27:443, protected by mutual TLS - and by an access control list that is wide open to anyone holding the client certificate shipped inside the malware. Each device is addressed by a ten-character token derived from a hash of its hardware fingerprint.

TopicDirectionPurpose
telemetry/devices/{id}/healthbot → C2Online/offline status, retained
telemetry/devices/{id}/reportsbot → C2Stolen data. Contacts and SMS are encrypted with the malware's own cipher; keystrokes are sent in the clear
telemetry/devices/{id}/streambot → C2Covert screen stream and accessibility tree
telemetry/devices/{id}/signalsC2 → botOperator commands to the device
frontend/updatesC2 → panelThe operator panel's event feed, in the clear
The live broker tree in MQTT Explorer, showing the telemetry and frontend branches and 22 connected clients
The live broker: Mosquitto 2.0.15, the telemetry/devices and frontend branches, and 22 clients connected at the time of observation

Capabilities, confirmed live

CapabilityDetail
SMS and OTP theftRead, send and hide SMS. Interception of one-time codes defeats two-factor authentication
Live keyloggerReal-time keystrokes - observed in WhatsApp, in Hebrew
Covert remote controlLive screen stream plus the accessibility tree, so the operator can drive the phone invisibly
Web-injectFake overlays drawn on top of banking apps
SpyingCamera, microphone, location, contacts, call log, accounts
SOCKS5 proxyFraud routed out through the victim's own mobile IP address
ResilienceAnti-uninstall, a watchdog, and automatic granting of Accessibility

What the stolen data looks like

The structures below are real. Every value that could identify a person has been replaced.

Beacon — in the clear

{"type":"ping","data":{"scr":"2","BC":"f","BP":"81",
  "permissions":{"accessibility":true,"ReadSMS":true,"SendSMS":true,
    "ReadContacts":true,"Camera":true,"RecordAudio":true,"Location":true,
    "Ant_delet":true,"HideSMS":false,"DefaultSmsApp":false}}}

Keystrokes — in the clear

{"type":"keys","data":{"data":
  "[12:04|com.whatsapp][REDACTED — Hebrew WhatsApp message][/com.whatsapp]"}}

SMS - encrypted on the wire, trivially decrypted

// on the wire:
{"type":"sms","data":"R2x1a0Zq9W... (ciphertext)"}

// after decryption — the stolen inbox, including bank one-time codes:
{"data":"[REDACTED sender] | [REDACTED — message body / one-time code] | 2026-11-15T.."}
The intercepted one-time code is exactly what lets the operator pass two-factor authentication. Everything else in this report is scaffolding around that single fact.

The victim profile, as the operator sees it

{"type":"device_update","device":{
  "deviceId":"[REDACTED]","phoneId":"[ANDROID_ID]",
  "manufacturer":"samsung","model":"SM-G990E","androidVersion":"Android 13",
  "ipAddress":"[REDACTED]","operator":"[carrier]","country":"il",
  "online":true,"injectionCount":18,"lockCredentialCount":3,
  "tag":"yad2","accessTag":"booking",
  "comment":"[REDACTED — Russian operator notes: screen PIN, credentials, bank-auth status]"
}}

The proxy - fraud from the victim's own address

A status message on the command channel confirms a live SOCKS5 back-connect proxy. The infected phone dials out to the operator's relay, traversing carrier NAT, and the relay exposes a per-victim listener.

{"type":"Socks5","cuz":"U","running":true,"connected":false,
 "username":"[REDACTED]","password":"[REDACTED]"}

Combined with covert screen control on the same device fingerprint and with stolen one-time codes, this makes a bank login originate from the victim's real device and their real IP address. The customer's entire trust context is reproduced, so geolocation, IP reputation, device binding and geo-velocity checks never fire. This is the part that should worry a fraud team most.


Covert remote control

In plain terms: the attacker can see the screen and tap it in real time, invisibly - as if holding the phone. That is how they operate the banking app themselves and get past SMS-based two-factor authentication.
  • Live screen pixels - a downscaled image of roughly 25 KB per frame, streamed continuously
  • The accessibility tree - every on-screen element with its text, its coordinates and whether it is clickable
  • Remote actuation - the operator taps any element by its coordinates, and so drives banking apps on the victim's own phone
A captured frame from a victim device showing the lock screen and a fake password prompt
A real captured frame: the victim's lock screen, with the malware's fake "update app password" prompt

The operator's command vocabulary

Across 14,928 command events on 34 devices over roughly ten hours, the complete observed vocabulary was remarkably small. Almost all traffic is an automated maintenance loop; the hands-on fraud commands surface only during a live session.

SignalCountDevicesMeaning
wrk · Apps·R·org.ypcsjwn.ldvhvd7,32130Uninstall a named package
wrk · Delete7,31630Trigger the uninstall and cleanup flow
com · sync28231Hourly inject-sync heartbeat
screen · Screen11Start the covert screen stream
NET_PING41Round-trip latency probe during live control
STOP_DIAGNOSTICS_STREAM41Stop the stream

The bot's own debug log shows every command routed through a modular dispatcher with eight registered handlers - Files, Camera, SMS, Activitys, Socks5, Permissions, Apps and Delete. That registry is the bot's full command surface. In this capture only three were exercised; the rest exist and were simply not used in the window.

One captured live session

On a single device we captured a complete operator screen-control session, start to stop, in about nine seconds.

20:29:43Z   screen   start stream   quality=28 width=280 fps=15   viewer_id=0ac958ad-…
20:29:43-49 NET_PING x4            latency probes during control
20:29:51-52 STOP_DIAGNOSTICS_STREAM  param_text="vnc" then "hvnc"   (same viewer_id)

The viewer_id identifies the operator's own console session - an attribution artefact that can link multiple devices and multiple sessions back to one operator.


The encryption is reversible

In plain terms: the cipher protecting the stolen data is very weak - a fixed, reversible XOR keystream, not AES. Anyone holding the sample can decode the traffic immediately, which is how we confirmed what is being taken.
out[i] = in[i] ^ ((3*i + K) & 0xFF) ^ TABLE[(K+i) % 32] ^ TABLE[(K+5*i) % 32]

TABLE = 7a3ca15e…5ca23e      K = 42      decrypt == encrypt

The device identifier is equally recoverable: a SHA-256 of the hardware fingerprint, reduced by a 32-bit hash and encoded into a ten-character alphabet with the ambiguous characters removed.


Attribution and scope

The operators are Russian-speaking - the profile comment field is Cyrillic and carries manual fraud notes. We observed 44 unique victims, 26 of them fully compromised, almost all Israeli across every major carrier, plus two in Poland. Campaign tags seen: yad2, tt18_octo, paradigm.

The victim table with device identifiers, addresses and operator notes redacted
The victim table. Device identifiers, addresses and operator notes are redacted; the scope columns are left visible

Indicators of compromise

TypeValue
APK SHA-256315d2e6acfbf1f745c53100b8909d5e7981c30a81e8ca2653036a55f937c7069
Signer SHA-256492682f877607ee99df2ddd2bd5953fd727bdf6e19d397de9dbbafd582bcad75
Packagesio.desk.crane.Hhopetomato · iacq.nom5w5vl.ucq58.* · org.ypcsjwn.ldvhvd
Native librarieslibyaan7ldva.so · libztx033yop.so
Command server194.5.97.27:443 — Mosquitto 2.0.15 over TLS
mTLS rootCN=AetherC2 Root CA · SHA-256 86266124958fae14046795179e8ca0ad94e1870f8b6b5cc2021d65d4207c24a2
Distribution URLhttps://yad2appl.org/app-download — fake Play page serving the APK
Operator sessionviewer_id 0ac958ad-… — console session identifier, useful for attribution

Detection and mitigation

  • Phishing. Block yad2appl.org. It exists only to serve a fake Play page that sideloads the APK.
  • Network. Block 194.5.97.27:443, and alert on outbound MQTT-over-TLS to a bare IP on 443 with no SNI from Android devices.
  • Endpoint. Flag any app that requests Accessibility and forty-plus permissions straight after a "Google Play update" screen, uses a silent session installer, or starts a local VPN service.
  • Fraud. Treat an Android device acting as a SOCKS5 exit node as a banking-fraud indicator, not merely a malware artefact. Alert on long-lived outbound tunnels, and flag sessions whose IP address and device both match the customer but whose behaviour does not.
  • Users. Never sideload a "delivery app" sent as a link after listing an item. Install from Google Play or not at all.
The hardest control to get right is the fraud one. Every other indicator here is a thing you can block; a login from the right device, the right IP address and the right one-time code is a thing you have to notice.

Key terms

TermWhat it means
Command server (C2)The attacker's server that infected phones connect to for orders and to upload stolen data
MQTTA lightweight messaging protocol — a chat bus. Here it is the command channel, hidden inside TLS on port 443
Covert remote controlSecret screen control: the attacker sees and taps the screen without the victim noticing
One-time code / 2FAThe code a bank sends by SMS. Stealing it defeats two-factor authentication
KeyloggerRecords everything the victim types, including passwords and messages
Web-inject / overlayA fake screen drawn on top of a real banking app to capture login details
Accessibility abuseAn Android feature built for users with disabilities, abused to auto-tap, read the screen and grant permissions
Mutual TLSBoth sides prove identity with certificates. The app carries the attacker's client certificate — which is how we got read access
Malware-as-a-ServiceThe kit is rented to operators, who each run their own victim campaigns

Found something that affects you, or want this checked against your own estate? Reach us at info@cyberapp.co.il.