"superior": A Botnet That Shipped Its Blueprint
The superior botnet hits routers, DVRs, and Android TV boxes with factory passwords. A leaked debug build gave up its blueprint — and then its full control protocol.
A botnet calling itself “superior” is actively compromising routers, digital video recorder (DVR) units, and Android TV boxes over telnet, logging in with factory-default passwords — nothing more. eFeSpain documented it in depth first. I’ve spent the last two weeks studying one of its builds using Anthropic’s Mythos 5.2, and the work turned up details that aren’t in anyone else’s public reporting yet — the build itself turned out to be the biggest finding. Here’s what the botnet does, why this build is unusual, and what businesses should do about the Android TV boxes in their offices.
What it does
Every observed session started over telnet with stock Internet of Things (IoT) credentials — admin/admin, root/root, root/GM8182, default/default. One honeypot recorded 559 admin/admin logins. After login, the attacker pastes a one-line shell command that downloads a dropper script from 77.239.124[.]167 and runs it. First contact came November 4–6. The main wave hit November 11: 753 telnet sessions from one loader address, 216 dropper runs, four dropper revisions in a single day. It was still active November 17.
The dropper is 12–14 kilobytes of commented shell, and it’s careful work. It reads the executable and linkable format (ELF) header of /bin/sh directly to pick the right central processing unit (CPU) architecture — some routers report one architecture while running another. It probes six directories for a spot where it can write and execute, since many boxes mount /tmp noexec. It downloads the matching binary using any of 17 fetch methods — curl, wget, busybox, trivial file transfer protocol (TFTP), python, even a hand-written hypertext transfer protocol (HTTP) request over netcat — checks for valid ELF magic before running anything.
The bot: encrypted command-and-control using X25519 key exchange and ChaCha20; a peer-to-peer mesh so commands relay node-to-node even when the main server is unreachable; persistence across reboots through cron jobs and init scripts — and on Android, a hidden Magisk/KernelSU module named media_svc; modules that kill rival malware, with operator-editable block and allow lists; and 36 attack methods in this build, including floods built to bypass Cloudflare and CAPTCHA protections, plus game-server and chat-platform attacks.
Why this one is unusual
One of the two ARM builds shipped with DWARF debug information intact — compiler metadata that maps a binary back to its source code. Most malware is stripped. This build exposed roughly 90 source module names, effectively its internal blueprint: hydra.c, swarm.c, scareware.c, bricker.c, kimwolf.c.
Then the coercion modules: an Android lock-screen trigger, framebuffer takeover (seizing the display feed itself), teletypewriter (TTY) scareware, a locker, and a bricker. The bricker code is present, but not confirmed to be reachable in practice.
What stood out to me is the telemetry. The operator runs a separate install-telemetry server that records each architecture reported, each binary fetched, and each launch — including whether the command-and-control port was reachable. Installation success, measured per device type. That’s operational discipline, applied to someone else’s hardware.
What our analysis adds
Details that aren’t in other public reporting as of November 17, per checks against public sources:
- The debug-build analysis and its ~90 module names
- The bricker module (code only)
- The telemetry server at 195.178.110[.]91:18081
- Two dropper revisions newly submitted to abuse.ch and VirusTotal
- The loader address 195.178.110[.]78 acting as a telnet loader
- YARA rules that match all six captured samples and none of the ~790 other files my honeypot collected in the same period
- Suricata network rules, syntax-checked with Suricata 7.0.10
Then the same debug build gave up the protocol. The command side of superior’s command-and-control channel isn’t in eFeSpain’s reporting. From static analysis I mapped it: the operator can send eight plaintext control commands — stop all attacks, self-destruct, push an update, pull a file, reconfigure its rival-killing module, hand out mesh peer lists, run a bandwidth benchmark on user datagram protocol (UDP) 5823, or request a self-report. Attack commands travel encrypted, Mirai-style, and this build registers 39 attack vector ids. The bot’s opening announcement is encrypted under a key baked into the sample, so bytes 2–5 of it are always identical — a reliable network signature. The repo includes an offline decoder that turns captured command-and-control and mesh traffic into readable JSON, plus a second Suricata ruleset covering that signature, the primary port (29444) and its fallback (14444), the plaintext status lines on transmission control protocol (TCP) 5821, a DuckDNS C2 domain now in the IOC list, and the mesh hello. Tested so far: the status-line and mesh-hello rules match the recorded lab traffic; the rest — including the announcement signature — are untested, because no registration was ever sent. Readable control traffic and a matchable handshake: detection an analyst can actually deploy.
The lab runs added a caveat worth stating plainly. In about ten minutes each, neither build ever looked up its C2 domain or connected to it — and in the debug build, nothing in the binary even opens the C2 socket. That protocol section is a reading of the code. These captured builds are driven over the mesh.
Then I ran it. Both builds went into virtual machines on an isolated network with no route out — a local fake internet answering every address, a silent sink accepting connections without replying. Self-replication, the killer, the locker and the cleaner were switched off with the bot’s own environment switches. Both builds came up the same way: a mesh listener on TCP 443 — raw TCP on the port every firewall assumes is HTTPS, not TLS — and SWR1 hellos toward the bootstrap address, plus thousands of plaintext status connections to port 5821 and exactly one geo lookup, ip-api.com, the only domain name either build ever queried. In the first run the two bots linked up, and all 37 hellos and frames in the capture decrypted and verified with keys derived from the fixed key in the sample: hello, peer exchange, ping, pong, and system facts, live on the wire. The mesh carries the same command vocabulary as the operator channel — attacks, scare, locker, kill — so these bots can drive each other with no server involved. The second run is its own lesson: the bots never linked, because neither was in the other’s ARP table, which is exactly how they find peers on a local network. eFeSpain had the mesh’s greeting and commands from the code, but in their captures it was switched off — never seen running until these lab runs.
The full write-up is in the research post, and everything — defanged indicators of compromise (IOCs), the YARA rules, both Suricata rulesets, the decoder — is in the GitHub repo under a Creative Commons Attribution (CC BY 4.0) license.
Method: static analysis — hashes, strings, ELF and DWARF metadata, disassembly — plus two isolated lab runs on November 17, in VMs on a network with no route out, with Mythos 5.2 assisting the analysis. No connection was ever made to attacker infrastructure; the samples ran only inside the lab.
The Android TV problem
Businesses are full of Android TV boxes: conference room displays, lobby signage, training rooms, warehouse dashboards. They run Android, often with Android Debug Bridge (ADB) network debugging left enabled, and they usually sit on the same network as workstations.
This is exactly what superior targets: ADB scanning, persistence through a fake Magisk module dressed up as a system service, and squatting on TCP port 5555 so rival bots cannot reuse the same foothold.
The consequences are plain. The device becomes distributed denial-of-service (DDoS) infrastructure. The scareware and locker modules mean screens can be locked or seized — visible disruption in a lobby or conference room. The bricker module, unconfirmed, signals appetite for permanent damage. And nothing observed required a zero-day. Factory passwords were enough, every time.
What to do now
- Disable telnet and change factory passwords on routers, DVRs, and cameras.
- Disable ADB on Android TV boxes, or limit it to USB.
- Put smart TVs and signage on an isolated network segment — a virtual local area network (VLAN).
- Check for unexpected listeners on TCP port 5555 and for Magisk/KernelSU modules nobody installed.
- Watch for raw, non-TLS traffic on TCP port 443 from appliances and TVs — this mesh runs there, and the repo’s Suricata rules match it.
- Grab the IOCs, plus the YARA and Suricata rules, from the GitHub repo.
The notable detail here is not the attack list. It’s the operational professionalism: per-architecture install metrics, launch tracking, a separate telemetry pipeline for its own deployment, commands that can retire a bot or push new code mid-campaign, a peer-to-peer mesh that keeps working with the C2 server gone — and a build that accidentally shipped its own blueprint. The defenses are boring, and they work: know what’s on your network, turn off the defaults, isolate the appliances.
Related: Our Managed Security services include continuous monitoring and vulnerability management for edge devices.

Chris Sewell
Chris Sewell
Senior Security & AI Solutions Architect | N2CON
Chris Sewell is a Senior Security & AI Solutions Architect at N2CON. He architects secure infrastructure, helps organizations responsibly integrate AI, and leads risk assessments a…