Skip to main content
Emerging ThreatsMalware & Ransomware

GitDoor Exposes Sophisticated Cloud Service Backdoors Targeting Asian and African Governments

Rows of computer servers and networking equipment in a brightly lit government agency server room.

"In early 2026, we identified a new, active intrusion campaign that we have named Operation ShadowCloud," the Securelist report said.

Initial infection: MeshCentral agent and sideloaded loaders

The operation begins with a MeshCentral (MeshAgent) installation observed inside a Taiwanese organization. Researchers recovered two MeshCentral Agent samples (MD5 74b025084026d8d9a64661bf995d835b) whose agent was hard‑coded to reach wss://27.102.66[.]162:443/agent.ashx. Once the WebSocket session was established, the attacker used the agent to drop three items on the endpoint: a legitimate executable used for DLL sideloading, a malicious loader DLL, and an encrypted file (config.ini) containing the core payload.

Two concrete instances were documented: one placing start.exe, sbiedll.dll and config.ini under C:\windows\system32\; another placing a Python binary and python39.dll under C:\Windows\system32\dts\pc\apo4x. The actors abused trusted binaries to implicitly load malicious DLLs — jli.dll, SbieDll.dll, python39.dll, WINMM.dll, and libcef.dll — next to legitimate executables whose MD5 hashes include javaw.exe (8b78ef56...), start.exe (be739a70...), pythonw.exe (84802caf...) and zoomit.exe (d955d031...).

Loaders and in‑memory execution techniques

Two distinct loader classes unpack and execute the encrypted payload entirely in memory. Type 1 loaders use a hybrid decode chain: a custom Base64 alphabet, an RC4 layer with a 16‑byte per‑sample key, then a second 16‑byte XOR; they even abuse RtlEthernetAddressToStringA and RtlEthernetStringToAddressA to transform decrypted bytes into MAC‑address strings and back before marking the memory PAGE_EXECUTE_READWRITE and jumping to an MZ header trampoline. Bootstrapping resolves APIs via CRC32 or, in a variant, via an FNV‑1 hashing routine and then reflectively loads the embedded backdoor (GraphDoor in one sample, MD5 14ec33ea...).

Type 2 loaders (SbieDll_st.dll, MD5 e63ec733...) take a simpler Base64→XOR path but add multiple runtime checks: they search kernel DLLs for a specific byte pattern, install a vectored exception handler, clear DR7 hardware breakpoints, verify the process name, and select native NT APIs (NtAllocateVirtualMemory, RtlMoveMemory, NtCreateThreadEx) on Windows 8 and newer or fall back to VirtualAlloc/CreateThread on older systems. Both loader types allocate memory, copy the decrypted backdoor bytes and start execution from the MZ header bootstrap.

GitDoor C2: GitHub issues, repository files and Ably channels

GitDoor (MD5 552f0a9cc5fc3755721fa49be05e8035) hides C2 traffic behind legitimate cloud services: GitHub repositories and the Ably messaging cloud. The backdoor decrypts repository URLs, a GitHub personal access token, HTTP headers and a User‑Agent string, then uses two repositories operated by the same GitHub account: Havo (download) and Trix (upload). In the analyzed samples those repositories resolved to https://github.com/LatonyCharnleyaW.

GitDoor creates a new GitHub issue in the Trix repo — Title “problem [GUID]”, Body “code [victim unique code]” — extracts the returned issue number and uses that unique code to tag files. It also connects to rest.ably.io using a Base64‑encoded Ably API key, and builds a filename main.%victim_unique_code%.js.map which it uploads to both Ably and GitHub. Victim system information is AES‑256‑ECB encrypted, Base64‑encoded, wrapped into JSON, then XOR‑encrypted with the filename and transmitted via POST to Ably and PUT to the GitHub repo.

The backdoor downloads plug‑ins (for example scr.ts and pr.ts) from the Havo repository. Those plug‑ins use a two‑stage decryption: Stage 1 AES‑256‑ECB with a hardcoded key "12345678901234567890123456789012", then a proprietary rolling XOR chaining scheme producing shellcode the backdoor maps and executes in memory. GitDoor then repeatedly labels/unlabels the GitHub issue (adding or deleting the "bug" label) and polls Havo’s contents to discover command files that match the victim’s unique code.

Custom shellcodes, modules and post‑compromise capabilities

Operation ShadowCloud delivers a modular set of COFF and in‑memory payloads. The campaign’s shellcodes implement multiple process‑injection techniques (CreateRemoteThread, SetThreadContext thread hijacking, NtQueueApcThread using ntdll syscalls, KernelCallbackTable manipulation) and provide an extensive command set: service enumeration and control, filesystem actions (recursive directory listing, secure wiping), network enumeration (netstat, ARP), file upload to upload.gofile[.]io, process dumping, and credential theft.

Notable modules include an imzimzgo 2.2.0 x64 Mimikatz build used with sekurlsa::logonpasswords full and lsadump::sam arguments, a WhatsAppKey extractor, a PetitPotam EFSRPC abuse implementation that constructs UNC paths and RPC bindings, and a WinPTY‑based interactive shell that drops winpty.dll and winpty‑agent.exe into %TEMP% and spawns cmd.exe with a fully interactive terminal. Command results are wrapped in a JSON object, AES‑encrypted, Base64‑encoded, XORed with a filename and uploaded to the attack repositories.

What this means for technologists, affected governments, and cloud providers

  • Technologists and security teams: look for MeshAgent activity with a hardcoded wss endpoint 27.102.66[.]162 and the MeshAgent MD5 74b025084026d8d9a64661bf995d835b; hunt for the loader MD5s (32151547..., e9d4081f..., e63ec733...) and for unexpected use of GitHub API calls that create issues, PUT content, or repeatedly add/remove labels tied to unusual filenames.
  • Affected governments and enterprises in Asia and Africa: the report identifies victims primarily within the government sector and documents one confirmed MeshCentral compromise inside a Taiwanese organization; breached networks should assume in‑memory backdoors and COFF shellcode families capable of credential theft, keylogging and proxying.
  • Cloud and platform operators: the campaign demonstrates how Ably and GitHub can be abused for C2 — the researchers enumerated multiple GitHub accounts and repositories created from late‑2024 through 2025 that the actor used as drop points and data stores.

Part 2 of the analysis will cover the GraphDoor backdoor, the plug‑in suite (keylogger, process listing, screen capture, a modified Erfrp proxy and an Ably channel proxy) and the recovered C2 addresses, as well as technical attribution notes connecting the operation to Chinese‑speaking threat actors, the report says.

Original report: https://securelist.com/operation-shadowcloud-gitdoor/122004/