Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
All full reports
Full reportCommodity malwareMonograph 44 / 5518 min read

IoT Botnet Infrastructure: Separating Kaiji Candidates from Router Peers

How 2,851 examined endpoints across five botnet families split into 1,696 withheld telecom peers, 195 commercial-VPS Kaiji candidates, and deployable network and Linux hunts

By
Odin by Brandefense
Published
19 November 2026
Reading time
18 min read
Edition
Version 1.1 · TLP:CLEAR
Scope
Prepared November 2026
IoT Botnet Infrastructure: Separating Kaiji Candidates from Router Peers

1. Executive Summary

In short

Odin by Brandefense reviewed an inventory of 2,851 endpoints on 2,587 hosts through 20 November 2026 across five botnet families: Mozi, Mirai, Hajime, Kaiji, and Gafgyt. Upstream feeds linked these families to 280,027 global sightings, and 1,831 examined records date before 6 November 2026. Nearly three in five endpoints (1,696, or 59.5%) sit on subscriber or regional telecom networks, where feed labels alone do not confirm infection. Zero endpoints met the confirmation bar for IP blocking. Enforce key-only SSH and keep telecom peer lists out of firewall blocks.

IoT and Linux DDoS botnets do not all run on the same infrastructure. Peer-to-peer router worms turn home and small-office gateways into temporary lookup nodes. Linux SSH botnets rent commercial virtual private servers (VPS) to stage binaries and send attack commands. Defenders can use cyber threat intelligence to review feed matches before permanent firewall blocks cut off broadband customers while missing internal peer traffic.

1

Mozi and Hajime peer-to-peer meshes have no central C2 server to block.

Mozi accounts for 215,742 feed-labelled sightings and 840 examined endpoints, with 84.3% sitting on dynamic telecom networks and zero dedicated C2 servers. Hajime adds 16,292 feed sightings and 837 examined endpoints, and all 20 feed-confirmed entries sit on Spanish fiber and Jordanian telecom networks as decentralized peers.

2

Regional and dynamic telecom networks explain all 66 feed-confirmed Mirai endpoints.

Out of 47,711 Mirai feed sightings, we examined 892 endpoints across 680 hosts. Dynamic broadband and mobile networks hold 598 of them, including three marked confirmed by upstream feeds. The other 63 of the 66 feed-confirmed endpoints sit on Ethiopian, Taiwanese, and Indian fixed or cable broadband carriers rather than commercial cloud servers.

3

Go-based Kaiji favors commercial VPS hosting over consumer broadband.

Unlike router swarms, Kaiji places 186 of its 195 examined endpoints outside dynamic subscriber networks and clusters 63.6% of them on ports 10081, 808, and 8080. One Kaiji endpoint on a commercial VPS remains uncorroborated as C2, as do all 87 Gafgyt entries.

Three Decisions for Leaders

  1. Keep uncorroborated feed lists off automatic firewall blocks. Zero endpoints in this set met the confirmation bar for permanent IP blocking. Route Mozi, Mirai, Hajime, Kaiji, and Gafgyt feed matches to network and host hunting instead of automated perimeter blocks.
  2. Isolate IoT segments and alert on peer-to-peer UDP traffic. Place IP cameras, routers, and building gateways on dedicated VLANs. Alert when any host on those segments sends BitTorrent Distributed Hash Table (DHT) queries or fetches BusyBox shell scripts.
  3. Require key-only SSH sign-in on every Linux server. Disable password login for root across cloud and on-premises Linux hosts. Monitor for binaries running from /usr/bin/lib/ and for SSH traffic on non-standard port 24.

The Study in Numbers

NumberWhat it is
280,027 / 2,851Global feed-labelled sightings across five botnet families, and the 2,851 endpoints we examined
2,587Distinct physical hosts behind the 2,851 examined endpoints, with zero hosts shared across families
1,696 (59.5%)Examined endpoints on dynamic subscriber ISPs (1,613) and regional telecom or carrier networks (83)
0 / 1,677Central C2 servers found across 840 Mozi and 837 Hajime peer-to-peer DHT endpoints
661 / 892Mirai endpoints on dynamic broadband (598) and regional fixed or cable telecom carriers (63)
0 / 195Confirmed Kaiji C2 addresses to block, with all 195 commercial-VPS and ISP Kaiji endpoints kept for hunting
7Published indicators in the Indicators tab: 7 JARM fingerprints to hunt (0 IPv4 addresses published)

How to use this report.

Security leaders can read Chapters 1, 6, and 7. Network and Linux teams should read Chapters 3, 4, 5, and 7. SOC and detection engineers should start with Chapters 5, 8, and 9. CTI analysts will find hosting and TLS detail in Chapters 2 through 6.

2. Four Infrastructure Groups, One Analytical Model

In short

None of the 2,587 hosts appears in more than one botnet family. We group the five families into four functional roles as an analytical model for defenders, not as a single operator.

Feed lists often mix decentralized router meshes, edge loader fleets, commercial Linux C2 servers, and benign port collisions into one flat table. Blocking every row treats a subscriber home router like a rented cloud server, cutting off the next broadband customer after DHCP lease rotation.

Comparing hosts across the 2,851-endpoint set shows zero shared hosts between Mozi, Mirai, Hajime, Kaiji, and Gafgyt. Shared TLS fingerprints appear across Mozi, Mirai, and Hajime; common JARMs alone do not confirm infection or establish the cause of TLS similarity. Figure 1 presents the four groups as an analytical model of infrastructure roles and defender actions.

Figure 1: P2P DHT peer meshes, router loaders, and Go SSH botnets split duties (analytical model)
Figure 1: P2P DHT peer meshes, router loaders, and Go SSH botnets split duties (analytical model)
GroupFamiliesExamined endpointsPhysical hostsPublished addressesDefender action
Peer-to-peer DHT meshesMozi, Hajime1,6771,6480Hunt UDP DHT and verify port 5555 on IoT VLANs
Edge router loader fleetsMirai, Gafgyt9797570Hunt /bins/ and BusyBox fetches; patch edge devices
Commercial-VPS Linux SSH botnetKaiji1951820Enforce key-only SSH; hunt ports 10081, 808, and 24
Port and certificate collisions (subset)Mozi, Mirai, Hajime, Kaiji12120Keep FTP, SSH, POP3S, and CDN origin certs out of blocks

The 12 collision and open-directory rows are counted inside the family totals above (1,677 + 979 + 195 = 2,851 endpoints across 1,648 + 757 + 182 = 2,587 physical hosts).

Where the endpoints sit

Across all 2,851 endpoints, China leads with 1,027 endpoints (36.0%), followed by India (140, 4.9%) and the United States (129, 4.5%). Russia (128, 4.5%), Indonesia (115, 4.0%), Taiwan (104, 3.6%), and the Philippines (98, 3.4%) come next. Brazil (68, 2.4%), Ethiopia (66, 2.3%), and Hong Kong (52, 1.8%) round out the top ten.

Network providerRoleEndpointsMain families
CHINA UNICOM China169 BackboneBroadband and mobile ISP833Mozi (639), Mirai (179), Hajime (10), Gafgyt (5)
CHINANET BACKBONEBroadband and mobile ISP129Mirai (124), Mozi and Hajime (5)
Globe Telecom Inc.Mobile and broadband ISP (Philippines)94Mirai (94)
Taiwan Fixed Network, Telco and Network Service Provider.Fixed-line telco (Taiwan)73Mirai (73)
EthioNet and Ethio TelecomNational telecom networks (Ethiopia)66Mirai (34 on EthioNet, 32 on Ethio Telecom)
PT Telekomunikasi IndonesiaNational broadband ISP (Indonesia)53Mirai (34), Hajime (19)
Spartan Host LtdCommercial VPS hosting (United States)38Kaiji (38)
NetLab GlobalCommercial VPS hosting (United States, Hong Kong)14Kaiji (14, including one feed-confirmed US VPS candidate and historical records for ports 24, 60000, and 60001 across three host tokens at different times)

Consumer broadband networks dominate Mozi, Mirai, and Hajime, while commercial VPS providers such as Spartan Host Ltd and NetLab Global appear almost exclusively under Kaiji.

3. Mozi and Hajime Peer-to-Peer DHT Swarms

In short

Mozi (840 examined endpoints) and Hajime (837 examined endpoints) use serverless BitTorrent DHT peer meshes, and feed labels on telecom endpoints do not by themselves confirm router infection. Neither family offers a central C2 address to block, so defenders must watch for bencoded UDP peer traffic and verify exposed ports instead.

Mozi and Hajime avoid central command servers by building on the BitTorrent Mainline DHT protocol (BEP 5). In public analyses of these families, peer routers, cameras, or digital video recorders join a Kademlia swarm over UDP (T1584.005) and relay signed configuration updates. Figure 2 maps all five families to their transport profiles and matching detections.

Figure 2: Protocol and port profiles separate P2P router peers from Kaiji VPS relays
Figure 2: Protocol and port profiles separate P2P router peers from Kaiji VPS relays

Mozi: 93.8% ephemeral high ports on subscriber broadband

Mozi accounts for 215,742 global feed sightings across 89 countries (77.0% of all sightings in this study). Across our 840 examined endpoints on 825 hosts, 788 endpoints (93.8%) use ports 10000 or higher across 712 distinct port numbers (T1571). Port 443 holds 38 endpoints, port 80 holds 7, and port 51769 holds 5.

Dynamic residential and mobile ISPs host 708 of the 840 Mozi endpoints (84.3%). China leads with 649 endpoints, including 639 on CHINA UNICOM China169 Backbone. India follows with 71 endpoints, including 34 on National Internet Backbone and 32 on Bharat Sanchar Nigam Ltd. Sweden (29), Pakistan (20), and Morocco (11) follow next. Seven Mozi endpoints on port 80 carried a confirmed label in upstream feeds on 13 November 2026, yet all seven sit on Pakistan Telecommuication company limited subscriber lines. All 825 Mozi host addresses are withheld.

On the wire, Mozi nodes send bencoded KRPC find_node queries (d1:ad2:id20:) and return signed config payloads carrying [count] and [ss] tags. On server or IoT VLANs where BitTorrent is prohibited, an outbound d1:ad2:id20: UDP packet flags DHT traffic for investigation; it does not by itself confirm infection.

Hajime: decentralized DHT peers and port 5555 listings

Hajime accounts for 16,292 global sightings across 149 countries and 837 examined endpoints on 823 hosts. Public analyses establish that Hajime has no central C2 server; its author pushes signed modules directly into the BitTorrent Mainline DHT swarm.

FeatureHajime count and share
Endpoints on ports 10000 and above649 of 837 (77.5%), across 743 distinct port numbers
Plain HTTP port 80 / TLS 443 / HTTP 808035 on 80, 3 on 443, 3 on 8080
Unverified TCP port 5555 listings3 Hajime-labelled records (port 5555 is commonly associated with ADB, but service banners were not captured)
Leading countriesRussia (116), Indonesia (71), India (58), Bangladesh (39), Ukraine (37), South Korea (35), Spain (33), Brazil (28)
Dynamic and carrier telecom endpoints291 dynamic ISP (34.8%) plus 20 carrier and telecom DHT peers

Upstream feeds marked 20 Hajime endpoints as confirmed on 13 and 14 November 2026. Nineteen listen on port 80 inside one Spanish subnet of LYNTIA NETWORKS S.A., a wholesale fiber carrier. The twentieth listens on port 443 on Batelco Jordan. It presents the same self-signed device certificate and JARM (04b06b00004b04b04b04b06b04b04be3b0c44298fc1c149afbf4c8996fb924) seen on two dynamic-ISP endpoints on Iran Telecommunication Company PJS. Because Hajime uses no central C2 and all 20 hosts belong to carrier or telecom networks, all 20 addresses are withheld.

4. Mirai and Gafgyt Edge Routers on Telecom Networks

In short

Mirai (892 examined endpoints) and Gafgyt (87 examined endpoints) spread BusyBox shell stagers across exposed routers and gateways. Of 66 Mirai endpoints marked confirmed by upstream feeds, 63 sit on regional fixed, cable, or fiber carriers and three sit on dynamic telecom networks. Isolate IoT VLANs and keep telecom peers off block lists.

Mirai and Gafgyt scan for exposed Telnet, SSH, and router web interfaces that defenders can find with external attack surface management (T1110.001, T1071.001). A shell one-liner invokes /bin/busybox with wget or tftp and chmod 777 to pull an architecture-specific ELF binary (/bins/*.mips, .mpsl, .arm7, .sh4). Figure 3 contrasts DHT gossip and BusyBox router staging with Kaiji's VPS-backed SSH workflow.

Figure 3: BitTorrent DHT gossip and Kaiji SSH staging diverge across network layers
Figure 3: BitTorrent DHT gossip and Kaiji SSH staging diverge across network layers

Why 66 feed-confirmed Mirai endpoints belong on watchlists, not block lists

Across 47,711 global Mirai sightings in 115 countries, we examined 892 endpoints on 680 hosts. High ports 10000 and above hold 762 endpoints (85.4%) across 714 distinct ports, while ports 443 and 80 hold 56 each. Leading countries are China (340), the Philippines (96), Taiwan (94), Ethiopia (66), Indonesia (42), Brazil (40), Albania (37), and South Africa (35).

Dynamic mobile and residential broadband backbones hold 598 Mirai endpoints (67.0%, including three marked confirmed by upstream feeds), and 231 more are uncorroborated or collision leads. Checking the six networks behind the other 63 of the 66 feed-confirmed Mirai endpoints (48 hosts) showed that all six are national fixed, cable, or fiber broadband carriers:

CountryTelecom or broadband networkEndpointsHostsPorts and TLS detail
EthiopiaEthioNet (AS24757)341918 on 443 and 16 on 80; all 18 TLS ports use the same CPE certificate seen on 4 Mozi-labelled endpoints, and 16 share JARM 21d19d00021d…
TaiwanTaiwan Fixed Network, Telco and Network Service Provider.1919All 19 on 443, sharing JARM 04b03b00004b…
TaiwanHoshin Multimedia Center Inc.66Plain HTTP port 80 on a cable broadband ISP
IndiaAtria Convergence Technologies Pvt. Ltd.,22Plain HTTP port 80 on ACT Fibernet consumer fiber
IndiaQuadrant Televentures Limited11Plain HTTP port 80 on a regional broadband carrier
IndiaRailTel Corporation of India Ltd.11Plain HTTP port 80 on a national broadband carrier

EthioNet is Ethiopia's national telecom carrier, sibling to the 32 Ethio Telecom rows already treated as subscriber broadband. Its JARM (21d19d00021d21d21c42d43d00000079c40eb3c1447424b28b311ae50d7d52) also appears on cable ISPs in Belize and Canada. Combined telecom networks account for 661 of the 892 Mirai endpoints (74.1%), and all 680 host addresses are withheld.

Gafgyt candidates

All 87 Gafgyt feed sightings span 16 countries across 77 hosts, with 64 listening on port 80 (73.6%). Seven sit on dynamic ISPs, while 80 remain uncorroborated candidates. Those candidates sit in the United States (18), the Netherlands (10), Cambodia (9 on OpenNet ISP Cambodia), and the United Kingdom (8). Their upstream feed record dates end in September 2026, before the November 2026 review cutoff, and all 77 host addresses are withheld.

5. Kaiji Golang SSH Botnet on Commercial VPS Hosting

In short

Kaiji is a Go-based Linux SSH brute-force and DDoS botnet whose infrastructure looks nothing like Mozi or Mirai. Commercial VPS providers host the bulk of its 195 endpoints, and 63.6% cluster on ports 10081, 808, and 8080. One upstream-confirmed NetLab Global endpoint lacked a second C2 signal and is withheld, so defenders should enforce key-only SSH and hunt Kaiji ports and host artifacts.

Written from scratch in Go, Kaiji guesses root SSH passwords on Linux servers (T1110.001, T1021.004). It copies disguised ELF binaries into /usr/bin/lib/ under names such as netstat or sshd and persists via /etc/init.d/. It also reads .ssh/known_hosts for lateral SSH movement and opens raw sockets for spoofed SYN_ACK, TCP, and UDP floods (T1498). Figure 4 compares Kaiji's VPS port records against shared router TLS profiles.

Figure 4: Commercial Kaiji VPS port records contrast with embedded router TLS stacks
Figure 4: Commercial Kaiji VPS port records contrast with embedded router TLS stacks

NetLab Global VPS candidates and structured port clusters

Across 195 Kaiji endpoints on 182 physical hosts in 17 countries, only 39 distinct ports appear and just 9 endpoints (4.6%) sit on dynamic ISPs. Commercial VPS hubs lead in the United States (73 endpoints, 37.4%), Hong Kong (39, 20.0%), and unattributed blocks (24, 12.3%). China (19, 9.7%), Japan (11, 5.6%), and Singapore (9, 4.6%) follow next:

Hosting providerKaiji endpointsDetail
Spartan Host Ltd38Largest commercial VPS cluster (19.5% of Kaiji endpoints)
NetLab Global141 feed-confirmed port-80 VPS candidate plus 3 host tokens with historical records for ports 24, 60000, and 60001 at different times
FranTech Solutions, Oracle, Cogent, Cloudflare London246 endpoints each across cloud and VPS hosting
CTG Server Limited and CNSERVERS LLC105 endpoints each on commercial server hosting

Three findings guide defense:

  • Upstream-confirmed VPS candidate (withheld). Recorded on 13 November 2026 on NetLab Global in the United States, one Kaiji-labelled commercial-VPS record lists port 80 (T1583.003). Because it showed no independent C2 behavior or second adversary signal beyond the upstream feed label, its address is withheld from block lists.
  • Core operator ports 10081, 808, and 8080. Three ports hold 124 of the 195 endpoints (63.6%): 10081 (59, 30.3%), 808 (49, 25.1%), and 8080 (16, 8.2%). Smaller clusters use 8888 (7), 60001 (6), 8087 (6), 23812 (5), 24 (4), 9527 (4), and 60000 (3). Ten Kaiji VPS endpoints share JARM 29d29d20d29d29d22c29d29d29d29d8e917db1de9d33002d70770504a55ef7.
  • Historical port 24, 60000, and 60001 records. Three Kaiji-labelled host tokens have historical records for ports 24, 60000, and 60001 at different times; the records do not establish a simultaneous port triad or bot control. Another NetLab Global host in Hong Kong has historical records for 8088 and 8089 at different times; they do not show both ports open together. Candidate addresses stay withheld for port and TLS hunting.

6. Feed Labels Against Confirmed C2 Infrastructure

In short

The 2,851 examined endpoints were drawn from 280,027 feed sightings; none yet meets the publication bar for an automatic IP block. Blocking raw IoT feed lists punishes broadband subscribers and breaks shared cloud services.

Figure 5 compares upstream feed-labelled records and examined endpoints with corroborated C2 block counts across all five families.

Figure 5: Feed-labelled IoT sightings reflect telecom peer records rather than C2 servers
Figure 5: Feed-labelled IoT sightings reflect telecom peer records rather than C2 servers
FamilyFeed sightingsExaminedPhysical hostsConfirmed C2 (block)Withheld peers, candidates, and collisions
Mozi215,7428408250840 (708 dynamic ISP, 36 suspected, 93 candidates, 3 collisions)
Mirai47,7118926800892 (598 dynamic ISP, 63 regional telco, 42 suspected, 185 candidates, 4 collisions or open dirs)
Hajime16,2928378230837 (291 dynamic ISP, 20 carrier or telco DHT peers, 17 suspected, 506 candidates, 3 collisions)
Kaiji1951951820195 (184 commercial VPS candidates, 9 dynamic ISP, 2 CDN collisions)
Gafgyt878777087 (80 historical candidates, 7 dynamic ISP)

Four reasons raw IoT botnet feeds misfire in firewalls

  1. Subscriber DHCP churn and regional carrier aliases. Dynamic ISPs hold 1,613 endpoints (56.6%), and regional carriers hold 63 more Mirai rows. Blocking ISP leases cuts off downstream broadband subscribers after reassignment.
  2. Serverless DHT peer protocols. Mozi and Hajime use signed BitTorrent DHT updates rather than central servers, so IP blocks on 20 Spanish or Jordanian peers cannot stop swarm propagation.
  3. Shared TLS fingerprints across families (120 endpoints across three JARMs). Three JARM hashes span Mozi, Mirai, and Hajime despite zero shared hosts. Hash 3fd3fd0003fd3fd00042d42d000000020120996177a65431cde640fa58d2e8 appears on 52 endpoints (32 Mozi, 14 Mirai, 6 Hajime). Hash 04d02d00004d04d04c04d02d04d04d9674c6b4e623ae36cc2d998e99e2262e appears on 44 endpoints (30 Mozi, 7 Mirai, 7 Hajime), and 29d14d00029d29d00029d14d29d29df26cd45df722aff084b72395f8d18d1b appears on 24 (18 Mirai, 5 Hajime, 1 Mozi). Common JARMs alone do not confirm infection, prove why TLS fingerprints match, or identify C2 software.
  4. Service-port, CDN, and open-directory collisions (12 endpoints). Collisions include FTP port 21 in Japan, SSH port 22 in Germany, and POP3S port 995 in China. Seven more match shared certificates (three CloudFlare Origin Certificate servers and four router vendor certs), while two Mirai candidates on 8080 and 8082 expose open HTTP directories.

7. What Defenders Should Do

Network and firewall

  • Monitor and restrict unexpected outbound traffic from Linux servers to primary Kaiji VPS operator ports (10081, 808, 8080, 8087, 8888) and relocated SSH port 24 at perimeter firewalls and proxies.
  • Block inbound WAN access to Telnet (23), Android Debug Bridge (5555), and router management ports (7547, 8080, 8443).

Linux servers and SSH

  • Enforce key-only SSH sign-in (PasswordAuthentication no and PermitRootLogin no) in /etc/ssh/sshd_config across all Linux servers (RFC 4252), restricted to jump hosts or VPNs.
  • Audit Linux hosts for binaries under /usr/bin/lib/, modified scripts in /etc/init.d/, and outbound SSH connections to port 24.

IoT and edge segments

  • Isolate IP cameras, DVRs, and branch SOHO routers on dedicated VLANs blocked from reaching internal server subnets.
  • Alert on outbound UDP packets to ports 10000 and above carrying BitTorrent Mainline DHT queries from server or IoT VLANs.

Detection and threat intelligence

  • Deploy the Suricata, Sentinel KQL, Sigma, and YARA rules in Chapter 8 with exact-hostname allowlists for dedicated lab or jump hosts.
  • Check autonomous-system ownership before blocking IoT feeds, keeping broadband ISPs, regional telcos, DHT peer lists, FTP/SSH/POP3S ports, and shared CDN certificates in hunting workflows.

8. Detection Rules

The full rules are in the Detection rules tab. During preparation, Suricata 7.0.12 (suricata -T) loaded all 7 signatures with 0 failures, and lint_suricata.py found 0 problems. The suite also passed YARA-X compilation (3 rules), pySigma validation (2 rules), and Kusto.Language semantic checks. Test and tune them against local traffic before automated isolation.

Rule fileWhat it catches
odin_iot_p2p_dht_mirai_and_kaiji_ssh_botnets.rulesSuricata 7: seven signatures for BitTorrent DHT queries, Mozi [count] responses, Mirai /bins/ and BusyBox stagers, ADB port 5555 shell commands, Kaiji .sh fetches, and SSH port 24
sentinel_iot_p2p_dht_and_kaiji_ssh_botnet_hunt.kqlMicrosoft Sentinel and Defender XDR: connections to Kaiji operator ports, high-port UDP peer fan-out, /usr/bin/lib/ binaries, and BusyBox stagers
sigma_linux_kaiji_ssh_persistence_and_iot_stager.ymlSigma (2 rules): Kaiji /usr/bin/lib/ execution or /etc/init.d/ staging, and BusyBox multi-architecture fetches
odin_kaiji_golang_ssh_mirai_and_mozi_dht_artifacts.yarYARA (3 rules): Golang Kaiji ELF implants, Mozi DHT config binaries, and Mirai or Gafgyt BusyBox shell droppers

How the Suricata signatures work

SIDLooks forWhy
9544001Outbound UDP packet starting with d1:ad2:id20: with 1:q and 1:y1:qe on ports 1024:65535BitTorrent Mainline DHT lookup from IoT or server VLANs
9544002Inbound UDP DHT response starting with d1:rd2:id20: and carrying [count]Mozi signed peer configuration sync payload
9544003Outbound HTTP GET for /bins/ ending in .arm, .arm5, .arm7, .mips, .mpsl, .x86, .spc, .sh4, .ppc, or .m68kMirai and Gafgyt multi-architecture ELF stager fetch
9544004HTTP request body containing /bin/busybox followed by chmodRouter command-injection exploit staging a binary
9544005Inbound TCP port 5555 packet carrying OPEN and shell:Unauthenticated ADB remote shell bootstrap
9544006Outbound HTTP GET for .sh on ports 808, 8080, 8087, 8888, or 10081 with Wget/, curl/, or Go-http-client/Kaiji shell stager fetch on primary VPS operator ports
9544007Outbound SSH 2.0 handshake on TCP port 24Relocated SSH port seen in Kaiji-labelled records

How the other rules work

The Sentinel KQL query correlates seven days of DeviceNetworkEvents and DeviceProcessEvents. It flags connections to Kaiji ports (24, 808, 8087, 10081, 60000, 60001) and high-port UDP fan-out to 25 or more public peers on ports 10000 and above. That network threshold measures high-port UDP fan-out rather than confirmed DHT payload traffic. The query also flags /usr/bin/lib/ execution and /bin/busybox stagers, with exact-hostname allowlists (CORP-SSH-BASTION-01, CORP-NETOPS-JUMP-01, IOT-FIRMWARE-LAB-01, SEC-SANDBOX-LNX-01) for known bastions and labs.

Sigma ruleLevelFires when
Kaiji Linux Botnet Masqueraded Binary Persistence in /usr/bin/libHighA process runs from /usr/bin/lib/ or stages there with chmod and /etc/init.d/ (exact lab hostnames excluded)
Mirai Mozi and Gafgyt BusyBox Multi-Architecture Stager ExecutionHigh/bin/busybox runs with wget http, tftp -g, or chmod 777 and /bins/, .mips, or .arm7
YARA ruleFires when
ODIN_Kaiji_Golang_Linux_SSH_DDoS_ImplantAn ELF under 15 MB contains Go build, /usr/bin/lib/ or /etc/init.d/, and .ssh/known_hosts or SYN_ACK
ODIN_Mozi_P2P_BitTorrent_DHT_Config_NodeA file under 2 MB combines d1:ad2:id20: or find_node with [count] or [ss] and /bin/busybox
ODIN_Mirai_Gafgyt_MultiArch_Shell_DropperA script under 512 KB combines /bin/busybox and chmod 777 with two architecture suffixes

MITRE ATT&CK mapping

TacticTechniqueHow it shows upDetection
Resource DevelopmentT1583.003 Acquire Infrastructure: Virtual Private ServerKaiji port records on commercial VPS hosts across ports 10081, 808, and 8080SID 9544006; KQL
Resource DevelopmentT1584.005 Compromise Infrastructure: BotnetBroadband and regional-telecom routers in Mozi, Hajime, and MiraiSIDs 9544001 to 9544005; KQL UDP fan-out
Credential AccessT1110.001 Brute Force: Password GuessingRoot SSH password guessing and router credential attacksKey-only SSH policy; SSH authentication logs
Lateral MovementT1021.004 Remote Services: SSHKaiji root SSH login, .ssh/known_hosts harvesting, and port 24 managementSID 9544007; Sigma rule 1; YARA
Command and ControlT1071.001 Application Layer Protocol: Web ProtocolsPlain HTTP binary staging on Kaiji operator ports and /bins/ pathsSIDs 9544003 and 9544006
Command and ControlT1571 Non-Standard PortEphemeral DHT ports >= 10000 and Kaiji ports 10081, 808, 24, 60000, 60001SIDs 9544001, 9544006, and 9544007; KQL
ImpactT1498 Network Denial of ServiceRaw-socket SYN_ACK, TCP, and UDP flood routinesStatic YARA implant indicator; no shipped rule detects flood traffic

9. Indicators

The Indicators tab holds 7 published indicators: 7 JARM fingerprints for TLS hunting. There are no published IPv4 addresses or file hashes in this report. Ten more protocol, path, port, and certificate markers appear in this report only.

TypeCountWhere to use it
JARM fingerprints7TLS hunting only; never block on a JARM alone
Wire, path, port, and certificate markers (text only)10Suricata, Sentinel KQL, Sigma, and YARA hunting

JARM fingerprints to hunt

JARMFamilyContext
29d29d20d29d29d22c29d29d29d29d8e917db1de9d33002d70770504a55ef7KaijiShared by 10 commercial-VPS Kaiji candidate endpoints
3fd3fd0003fd3fd00042d42d000000020120996177a65431cde640fa58d2e8Mozi, Mirai, HajimeShared TLS fingerprint across 52 feed-labelled endpoints (32 Mozi, 14 Mirai, 6 Hajime)
04d02d00004d04d04c04d02d04d04d9674c6b4e623ae36cc2d998e99e2262eMozi, Mirai, HajimeShared TLS fingerprint across 44 feed-labelled endpoints (30 Mozi, 7 Mirai, 7 Hajime)
29d14d00029d29d00029d14d29d29df26cd45df722aff084b72395f8d18d1bMirai, Hajime, MoziShared TLS fingerprint across 24 feed-labelled endpoints (18 Mirai, 5 Hajime, 1 Mozi)
21d19d00021d21d21c42d43d00000079c40eb3c1447424b28b311ae50d7d52MiraiTelecom CPE self-signed TLS stack across 21 Mirai-labelled endpoints
04b03b00004b04b04b04b03b04b04b9674c6b4e623ae36cc2d998e99e2262eMiraiFixed-telco gateway TLS cluster across 19 endpoints in Taiwan
3fd3fd16d3fd3fd22c3fd3fd3fd3fdf20014c17cd0943e6d9e2fb9cd59862bHajimeLeading TLS fingerprint across 12 Hajime-labelled edge peer endpoints

Other indicators (text only)

IndicatorTypeWhere to hunt
d1:ad2:id20:Bencoded BEP 5 KRPC node querySuricata SID 9544001 on IoT and server VLANs
[count]Mozi signed DHT config sync tagSuricata SID 9544002 and YARA
/bins/Multi-architecture ELF HTTP stager URI prefixSuricata SID 9544003 and Sigma rule 2
/bin/busyboxEmbedded shell dropper command markerSuricata SID 9544004, KQL, Sigma rule 2, and YARA
OPEN + shell: on TCP 5555Unauthenticated Android Debug Bridge shell bootstrapSuricata SID 9544005
/usr/bin/lib/Kaiji masqueraded Linux persistence directorySentinel KQL, Sigma rule 1, and YARA
TCP 10081, 808, 8080Primary Kaiji VPS listener ports (124 of 195 endpoints)Suricata SID 9544006 and Sentinel KQL
TCP 24, 60000, 60001Historical Kaiji-labelled ports recorded at different times across three host tokensSuricata SID 9544007 and Sentinel KQL
CN=localhostDefault self-signed TLS subject (6 Mirai, 4 Kaiji)TLS hunting context only
CN=CloudFlare Origin CertificateShared CDN origin certificate on 3 collision endpointsExclude from standalone certificate blocks

Withheld: all 2,587 physical host addresses across the study, plus hardware-identifying device certificate strings. The withheld set includes 1,613 dynamic subscriber ISP rows and 68 hosts behind 83 regional-telecom and carrier DHT peer rows. It also includes all 182 Kaiji hosts, other uncorroborated candidates, and collision or open-directory hosts.

10. Scope and Limits

  • Window. The inventory was reviewed through 20 November 2026, but its upstream records span earlier years: 1,831 of 2,851 were last seen before 6 November, including 189 Kaiji rows and all 87 Gafgyt rows.
  • Feed labels are not attribution. Family names (Mozi, Mirai, Hajime, Kaiji, Gafgyt) come from upstream feeds. Zero endpoints in this dataset met the multi-signal confirmation bar for published adversary C2 infrastructure.
  • The model is an analytical model. Zero hosts overlap across the five families. Figures 1 and 3 show how router peer worms and VPS-hosted SSH botnets divide roles, not one coordinated actor.
  • Telecom and subscriber hosts are third parties. Consumer broadband gateways, regional telecom CPE hosts, and DHT peer records belong to subscribers and carriers, and feed labels alone do not confirm infection, so their addresses and device strings are withheld.
  • No downstream flood visibility. External records do not show which targets a botnet attacked or capture internal packet logs from target Linux servers.
  • VPS address rotation. Commercial VPS addresses change hands when leases expire, so uncorroborated single-feed VPS listings should stay in hunting workflows rather than permanent firewall block lists.
  • Rule validation. All four rule files passed syntax and semantic checks on local Suricata 7.0.12, YARA-X, pySigma, and Kusto.Language tooling, without live traffic replay. Validate them against local baselines before enabling automated isolation.

11. Sources

  1. MITRE ATT&CK. T1584.005 Compromise Infrastructure: Botnet. https://attack.mitre.org/techniques/T1584/005/ (verified 21 November 2026)
  2. BitTorrent.org. BEP 5: DHT Protocol Specification. https://www.bittorrent.org/beps/bep_0005.html (verified 21 November 2026)
  3. MITRE ATT&CK. T1583.003 Acquire Infrastructure: Virtual Private Server. https://attack.mitre.org/techniques/T1583/003/ (verified 21 November 2026)
  4. MITRE ATT&CK. T1110.001 Brute Force: Password Guessing. https://attack.mitre.org/techniques/T1110/001/ (verified 21 November 2026)
  5. MITRE ATT&CK. T1021.004 Remote Services: SSH. https://attack.mitre.org/techniques/T1021/004/ (verified 21 November 2026)
  6. IETF. RFC 4252: The Secure Shell (SSH) Authentication Protocol. https://www.rfc-editor.org/rfc/rfc4252 (verified 21 November 2026)
  7. MITRE ATT&CK. T1571 Non-Standard Port. https://attack.mitre.org/techniques/T1571/ (verified 21 November 2026)
  8. MITRE ATT&CK. T1071.001 Application Layer Protocol: Web Protocols. https://attack.mitre.org/techniques/T1071/001/ (verified 21 November 2026)
  9. MITRE ATT&CK. T1498 Network Denial of Service. https://attack.mitre.org/techniques/T1498/ (verified 21 November 2026)
  10. Salesforce Engineering. JARM: An Active Transport Layer Security (TLS) Server Fingerprinting Tool. https://github.com/salesforce/jarm (verified 21 November 2026)

What is…?

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

All explanations

Frequently asked questions

Why do Mozi and Hajime have zero central C2 servers to block?

Mozi (215,742 feed sightings; 840 examined endpoints) and Hajime (16,292 sightings; 837 examined endpoints) do not rely on a central command server. Both families build serverless peer-to-peer meshes on top of the BitTorrent Mainline Distributed Hash Table (DHT) protocol. In public analyses of both families, each node joins a UDP swarm, looks up peers with d1:ad2:id20: queries, and relays signed configuration updates.

Why did 66 feed-confirmed Mirai endpoints stay off the block list?

Mirai accounts for 47,711 global sightings (892 examined endpoints on 680 hosts) and Gafgyt accounts for 87 sightings (87 examined endpoints on 77 hosts). Both families spread across routers and cameras via Telnet brute-force attacks and /bin/busybox shell one-liners that fetch /bins/ ELF stagers.

How does the Go-based Kaiji Linux SSH botnet operate?

Kaiji was written from scratch in Go to brute-force root SSH logins on Linux servers. It hides binaries in /usr/bin/lib/, hooks /etc/init.d/, reads .ssh/known_hosts for lateral movement, and launches raw-socket DDoS floods.

Defender kit

Indicators of compromise

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

IDTypeIndicatorPortContextAction
R44-IOC-01 JARM 29d29d20d29d29d22c29d29d29d29d8e917db1de9d33002d70770504a55ef7 808 kaiji · shared by ten commercial VPS endpoints Hunt
R44-IOC-02 JARM 3fd3fd0003fd3fd00042d42d000000020120996177a65431cde640fa58d2e8 443 mozi, mirai and hajime · shared TLS fingerprint across 52 feed-labelled endpoints Hunt
R44-IOC-03 JARM 04d02d00004d04d04c04d02d04d04d9674c6b4e623ae36cc2d998e99e2262e 443 mozi, mirai and hajime · shared TLS fingerprint across 44 feed-labelled endpoints Hunt
R44-IOC-04 JARM 29d14d00029d29d00029d14d29d29df26cd45df722aff084b72395f8d18d1b 443 mirai, hajime and mozi · shared TLS fingerprint across 24 feed-labelled endpoints Hunt
R44-IOC-05 JARM 21d19d00021d21d21c42d43d00000079c40eb3c1447424b28b311ae50d7d52 443 mirai · telecom CPE self-signed TLS stack across 21 endpoints Hunt
R44-IOC-06 JARM 04b03b00004b04b04b04b03b04b04b9674c6b4e623ae36cc2d998e99e2262e 443 mirai · fixed-telco gateway TLS cluster across 19 endpoints in Taiwan Hunt
R44-IOC-07 JARM 3fd3fd16d3fd3fd22c3fd3fd3fd3fdf20014c17cd0943e6d9e2fb9cd59862b 443 hajime · leading TLS fingerprint across 12 edge peer endpoints Hunt

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_iot_p2p_dht_mirai_and_kaiji_ssh_botnets.rules 13 lines
Download
YARA odin_kaiji_golang_ssh_mirai_and_mozi_dht_artifacts.yar 62 lines
Download
Sentinel KQL sentinel_iot_p2p_dht_and_kaiji_ssh_botnet_hunt.kql 68 lines
Download
Sigma sigma_linux_kaiji_ssh_persistence_and_iot_stager.yml 79 lines
Download