Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
All full reports
Full reportCredential theftMonograph 06 / 1923 min read

Keystrokes Before Submit: AiTM Workbenches, noVNC and Ligolo-ng Tunnels

What 2,750 recently active phishing, proxy and tunnel listeners and three exposed workbenches show about post-MFA intrusion

By
Odin by Brandefense
Published
19 November 2026
Reading time
23 min read
Edition
Version 1.1 · TLP:CLEAR
Scope
Prepared November 2026

1. Executive Summary

In short

In a two-week window, with data up to 16 November 2026, Odin by Brandefense reviewed 2,750 recently active phishing, proxy and tunnel listeners and three exposed workbenches. One server holds a kit built to capture keystrokes before a login form is submitted, beside Evilginx2 and Cobalt Strike. Another pairs a Mitmproxy listener with noVNC client files. Ligolo-ng relays favour port 11601 and HTTPS. These are separate findings, not one campaign. Require origin-bound passkeys, block outbound ports 11601 to 11610, and keep vendor proxies off blocklists.

Attackers who want to get past multifactor authentication (MFA) use three kinds of tools. Phishing managers such as GoPhish send the lures. Adversary-in-the-middle (AiTM) proxies such as Evilginx2 and Mitmproxy sit between the user and the real login page and capture session cookies. Tunnel relays such as Ligolo-ng then carry the attacker into internal networks. This report looks at the infrastructure behind each stage, and at three exposed operator workbenches.

1

A cloned login kit is built to send keystrokes before the user submits.

On Host A in the Philippines (address withheld, because a red-team exercise cannot be excluded), a 247-file open directory holds a static login cloner and a seven-line keystroke script. The script is designed to send typed text to the server when the user changes field or closes the tab. Evilginx2 and Cobalt Strike directories sit on the same server. Who operates it, and whether it is an authorized red-team host, is unknown.

2

Two unlinked workbenches hold M365 enumeration scripts and noVNC client files.

Host C in Indonesia (832 files) holds scripts written to check whether Microsoft 365 accounts exist and to test IMAP logins. Host B in India (117 files) pairs a confirmed Mitmproxy listener with a noVNC browser-viewer client. Replaying cookies from Host B would also need a browser, a VNC server and a WebSocket bridge. None was observed.

3

Recently active Ligolo-ng relays favour default port 11601 and HTTPS 443.

Of 347 recently active Ligolo-ng relays on 310 IPs in 42 countries, 37.2% listen on TCP 11601 and 27.1% on TCP 443. Global records hold 938 Ligolo-ng servers in 54 countries. Eleven relays carry third-party feed labels that link them to YoroTrooper-tagged infrastructure.

Three Decisions for Leaders

  1. Require origin-bound passkeys at sign-in. W3C WebAuthn passkeys with device-compliance Conditional Access stop reverse-proxy and remote-viewer lures at the login step. Add Entra ID Token Protection and DBSC pilots for protection after sign-in.
  2. Block default Ligolo-ng ports and hunt non-browser tunnels. Deny outbound TCP 11601 to 11610. Correlate risky Entra ID sign-ins with non-browser TLS or WebSocket connections from the same user within 24 hours.
  3. Keep blocklists for corroborated hosts only. Commercial proxy networks, awareness-training platforms, appliance certificates and stock open-source file hashes do not belong on blocklists. Keep single-banner listeners and dual-use workspaces on watchlists until a lure or egress signal appears.

The Study in Numbers

NumberWhat it is
2,750Recently active phishing, proxy and tunnel listeners on 2,212 IPv4 hosts in 69 countries
26,329Global banner records for GoPhish (23,207), Mitmproxy (1,877), Ligolo-ng (938) and EvilGoPhish (307)
48Listeners with commercial proxy (44) or awareness-training (4) certificate names, held out of blocklists
136,742 bytesSize of the log file beside the keystroke receiver on Host A (1,310 lines, contents not established)
1,196Files on three exposed workbenches: 247, 117 (Host B) and 832 (Host C)
347Recently active Ligolo-ng relays on 310 IPs in 42 countries
3Published indicators: custom script hashes from Host A (selected set of 28; addresses withheld)

How to use this report.

Security and identity leaders can read Chapters 1, 4 and 7. SOC engineers and threat hunters should start with Chapters 3, 5, 8 and 9. CTI analysts will find the infrastructure detail in Chapters 2 to 6.

2. The Landscape: Phishing, Proxy and Tunnel Listeners

In short

Most phishing and proxy listeners on the internet are unconfirmed. Among those that are confirmed, 48 carry the certificate names of a commercial proxy network or an awareness-training vendor, so raw listener feeds are not safe to block.

Three stages, three tool families

As an analytical model, the tools in this report map to three stages of a post-MFA intrusion. The model describes tool roles; it does not claim one actor uses all three.

StageToolsWhat they doMITRE ATT&CK
Lure deliveryGoPhish, EvilGoPhishSend phishing campaigns and track clicks(delivery)
Credential and session interceptionEvilginx2, MitmproxyProxy the real login page and capture cookiesT1557, T1539
Network pivotingLigolo-ngTunnel Layer-3 traffic into internal subnetsT1572

Global records and recent activity

Global records count every banner on file for each framework. The recently active set is the smaller group that answered within the observation window. It includes 1,200 GoPhish and 1,200 Mitmproxy listeners (a capped selection), all 347 recently active Ligolo-ng relays and 3 FRP control panels as a baseline.

FrameworkGlobal records (countries)Recently active (IPs / countries)Main ports in the active setTop hosting providers in the active set
GoPhish23,207 (119)1,200 (1,080 / 62)3333 (764, 63.7%), 443 (142), 80 (32), 8443 (31), 3000 (16), 21 (11)DigitalOcean (137), Amazon.com (127), OVH SAS (112), Microsoft (73), Google (54)
Mitmproxy1,877 (46)1,200 (820 / 38)32300 (304, 25.3%), 8081 (102), 8080 (80), 8443 (57), 8888 (19), 443 (19)ALFA TELECOM (190), Hangzhou Alibaba (145), Serverel (114), Hetzner (55), DigitalOcean (49)
Ligolo-ng938 (54)347 (310 / 42)11601 (129, 37.2%), 443 (94, 27.1%), 80 (20), 8443 (14), 8080 (14), 2083 (12)OVH SAS (33), Microsoft (28), DigitalOcean (26), UpCloud (22), Amazon.com (21), Contabo (21)
EvilGoPhish307 (32)Not in the active set443, 8443, 3333 (local API)Cloud VPS; source found on Host A
FRP (baseline)n/a3 (3 / 3)7500 (3)Snowd Security, Hangzhou Alibaba, Unesty (1 each)

The 2,750 listeners sit on 2,212 IPv4 hosts in 69 countries and 462 autonomous systems.

Figure 1: Certificate checks set aside 48 vendor collisions from 2,750 listeners
Figure 1: Certificate checks set aside 48 vendor collisions from 2,750 listeners

How the 2,750 listeners break down

"Framework confirmed" names the software. It does not mean hostile intent.

StatusListenersWhat it containsHandling
Watchlist, framework confirmed193110 Mitmproxy, 72 GoPhish web or admin listeners, 11 FTPS rows reusing a GoPhish certificateWatch; block only after a lure or egress signal
Watchlist, vendor certificate name4844 commercial proxy and 4 awareness-training certificate names, chain trust not verifiedWatch; verify the certificate chain before excluding or blocking
Suspected5046 GoPhish, 1 Mitmproxy, 3 Ligolo-ngSend to threat hunting
Candidate2,4521,067 GoPhish, 1,043 Mitmproxy, 339 Ligolo-ng, 3 FRPHold; do not block
Held out72 FortiGate appliance certificates, 5 hits in a U.S. public-sector networkKeep off blocklists

Figure 1 counts the 7 held-out rows inside their software status, so its legend shows 195 confirmed, 52 suspected and 2,455 candidates.

GoPhish: port 3333 and shared certificates

Of the 1,200 GoPhish listeners, 764 (63.7%) expose the default admin port 3333. Of those, 756 are candidates and 8 are confirmed.

Four confirmed rows carry the certificate name of a commercial security-awareness training vendor. Ownership and certificate-chain trust were not verified, so they are held out as possible collisions. That leaves 83 confirmed GoPhish rows: 72 web or admin listeners plus 11 FTPS rows.

The 72 web or admin listeners run on 443 (30), 8443 (12), 3000 (11), 3333 (8) and other web ports (11). OVH SAS hosts 46 of the 83 rows (55.4%), followed by Amazon.com (6), DigitalOcean (5) and netcup GmbH (1).

The 11 rows on TCP port 21 are FTPS daemons on OVH SAS. They reuse the same self-signed provider-default certificate that GoPhish uses on 443 and 8443 on the same cloned VPS hosts (33 rows share it). This is certificate reuse across services, not GoPhish speaking FTP.

Mitmproxy: a commercial proxy collision

Mitmproxy candidates cluster on high proxy ports: 32300 (304, on ALFA TELECOM s.r.o. and Serverel Inc.), 8081 (102) and 8080 (78 candidates plus 2 confirmed).

Of 156 confirmed Mitmproxy rows, 44 (28.2%) carry certificate names that match a commercial proxy network. Chain trust was not verified, so they stay on the watchlist as possible collisions. They make up all 37 of DigitalOcean's confirmed Mitmproxy rows and 7 of 8 on Constant/Vultr.

Removing the 44 leaves 112 confirmed rows, or 110 after setting aside 2 AWS listeners with CN=FortiGate appliance certificates. Hangzhou Alibaba Advertising hosts 65 of the 112 (58.0%). Of these, 64 share two private-range certificate names, 172.29.97.219 and 172.18.9.199, on 6xxx ports. Smaller groups sit on CV Griya Cloud Server (7), Contabo (7), Hetzner (6), AROSSCLOUD (1), HostPapa (1), Psychz Networks (1) and the small provider hosting Host B.

3. A Keystroke Streamer Beside Evilginx2 and Cobalt Strike

In short

An open directory on Host A holds three toolsets on one server: a static login cloner with a keystroke script, an Evilginx2 installation, and Cobalt Strike files. A 136,742-byte log file sits beside the keystroke receiver; its contents are not established.

Why capture keystrokes before submit

A static phishing clone only receives a password when the user clicks Submit. If the user types part of a password, notices the wrong domain and closes the tab, a normal clone gets nothing. The kit described here is designed to close that gap.

On 16 November 2026, Odin by Brandefense observed an open directory on Host A (Philippines). It exposes 247 files across 591 directories (4,258,037 bytes). Figure 2 maps how the parts fit together.

Figure 2: Pre-submit sendBeacon keylogging co-resides with Evilginx2 on Host A
Figure 2: Pre-submit sendBeacon keylogging co-resides with Evilginx2 on Host A

The static cloner kit

The kit has three files under /phish/. The table describes each one by structure only. No script content is reproduced.

FileSize and linesRole
switch_site.sh3,063 bytes, 87 linesBash cloner. Downloads a target login page and its assets with a Chrome/133.0.0.0 User-Agent, inserts a script tag that loads k.js, and writes the real login URL to redirect.txt (51 bytes)
k.js811 bytes, 7 linesKeystroke script. Labels each input field by type, name or id, marks deletions with a text token, and sends the buffer to the server path /k
receiver.py3,350 bytes, 89 linesPython web server on port 8080. Serves the clone and the script, accepts posts to /k, sets a 5-second socket timeout and a wildcard CORS header

The keystroke script uses the browser Beacon API (sendBeacon). A browser queues a Beacon request in the background and sends it even while the tab is closing. The request is a small text/plain POST. The script falls back to a keep-alive fetch only when the Beacon API is missing.

The script sends its buffer at two moments: when the user moves to a different field, and when the page unloads. That is why the kit can collect partial input from a user who never submits.

The cloner fetches every asset path without logging in. On single-page apps, this leaves a trace. Three files on the server with image names hold the same 27,793-byte HTML landing page instead of an image.

receiver.py carries comments in Simplified Chinese about protection against slow sockets. That is an artifact of whoever wrote the toolkit. It is not attribution of the operator.

Evilginx2 and Cobalt Strike on the same server

The same open directory holds a full Evilginx2 tree and a Cobalt Strike workspace. Paths and secrets are withheld.

ComponentWhat is exposed
Evilginx2 source/evilginx2/ checkout built with Go 1.22. Its CHANGELOG opens with # Unreleased above # 3.3.0, so it is a post-3.3.0 master branch. Stock core/phishlet.go (29,710 bytes) and core/gophish.go (2,988 bytes)
Evilginx2 runtimeA BuntDB session store (21,293 bytes, 196 lines), a private root CA key file (1,679 bytes), a visitor blacklist with 21 addresses (withheld), and a config file (4,897 bytes)
Evilginx2 configurationBinds its own public address (withheld) as external_ipv4, sets blacklist.mode: off, and lists two active lure paths
PhishletsTwo custom YAML phishlets (3,275 and 5,811 bytes) declare author: 'redteam' and target two Chinese consumer portals. Neither injects the keystroke script
Cobalt StrikeA keystore, teamserver authorization files, a beacon key pair, a credential store, the launcher source-common.sh, and TightVNC post-exploitation DLLs winvnc.x64.dll and winvnc.x86.dll

The static cloner and Evilginx2 are co-located, not integrated. The phishlets do not load k.js. The author: 'redteam' tag and the unknown authorization mean a red-team exercise cannot be excluded.

File hashes for hunting

Hashes were recorded at acquisition.

FileSizeSHA-256Use
/phish/k.js811 bytesdc89160b8501065099a69bc6a8f34ce085774878387d071f4e249eebb57a757bBlock: custom keystroke script
/phish/receiver.py3,350 bytes629fd5033240e77684003b632c3b247671ac2043a4e87ae7f420feb836551a45Block: custom clone server and /k receiver
/phish/switch_site.sh3,063 bytes7da269d7b5b55d409b71702dc53d4329bbeb0afb0540ab3d98896c3625836c54Block: custom portal cloner
/evilginx2/core/phishlet.go29,710 bytes00d134e5be68c83a13c35101cd26aca6efe6af608144c35add5edf3a4139f336Context only: stock upstream source
/evilginx2/core/gophish.go2,988 bytes6097864f26657ac5add013af5e17f508678dec3eb57548dbf37ea09e2a02a78cContext only: stock upstream source

The two Evilginx2 source files are unmodified upstream code. Do not block them on their own.

4. M365 Enumeration Scripts and a Co-Hosted noVNC Client

In short

Two separate, unlinked servers sit on the watchlist. Host C holds scripts written to enumerate Microsoft 365 accounts and test logins. Host B pairs a confirmed Mitmproxy listener with noVNC client files, which are one part of a possible replay setup.

Why replay location matters

When an attacker steals a session cookie through a cloud proxy and replays it from a laptop or VPN, identity systems often notice. Continuous Access Evaluation and risk-based Conditional Access react to the change of IP, network and location.

One way around this is to replay the cookie from the proxy server itself, through a remote browser. Host B shows one part of that setup. Figure 3 shows what was seen and what would still be needed.

Figure 3: Co-hosted noVNC files are one part of a replay setup while scripts probe M365
Figure 3: Co-hosted noVNC files are one part of a replay setup while scripts probe M365

Host C: an AI pentest workspace with M365 recon scripts

On 13 November 2026, Host C in Indonesia exposed 832 files (1,280,209 bytes). The directory was created by an open-source AI pentest workspace. That software is dual-use, and the scripts carry diagnostic strings such as zz-diag and auth-test. No result on any target is verified, so Host C stays on the watchlist and its address is withheld.

FileSize and linesSHA-256What it was written to do
/o365_test.py1,719 bytes, 2624146b8cbbafc57995e77c7a65e799fc471374b27c06a7ac931a7282c2c9cfaaQuery M365 GetCredentialType for 4 accounts and test IMAP login on port 993
/mp_login.html12,573 bytes, 109Withheld (copied third-party page)Cloned merchant billing login page (also at /mplogin.html)
/captcha_probe.py2,212 bytes, 51e0f5c20fa788887402f33aefbd98b5e6a5d33c2dc5af529adf72dcea7bbca871Test null and array CAPTCHA parameters and Tesseract OCR thresholds
/lateral_auth.py2,153 bytes, 35e5c7d66e25771dded8a94810bd4ffe25ca05af325c30c666e29543a83d606f02Test credential reuse across SMTPS, Zimbra and Jenkins

Account enumeration. o365_test.py is titled as an account-existence check. It was written to call the public login.microsoftonline[.]com/common/GetCredentialType endpoint and print whether each account exists, its throttle status, domain type and preferred credential. The isFidoSupported flag in its request is a standard client field, not a tenant check. A second loop was written to test basic-auth IMAP against outlook.office365[.]com:993. Outcomes are not verified.

Lures and CAPTCHA tests.

No Microsoft 365 phishing page sits on Host C. The lures are merchant billing login clones and a dark-theme Jenkins login (/login.html, 2,337 bytes). captcha_probe.py was written to test edge-case CAPTCHA inputs against a merchant endpoint.

Credential reuse and scheduler scripts.

lateral_auth.py was written to test reuse on SMTPS port 465, the Zimbra preauth service on 8443 and the Jenkins login check. Two more scripts, cr_pwn.py and cr_rce_proof.py, were written to modify jobs on a merchant's Cronicle scheduler with a callback to Host C. Target-side outcomes are not verified.

Host B: Mitmproxy beside a noVNC client

Host B is a VPS at a small hosting provider in India. It runs a confirmed Mitmproxy TLS listener. On 13 November 2026, the same IPv4 address also served an open HTTP directory of 117 files (917,101 bytes). It is a complete noVNC 1.x HTML5 remote-framebuffer client, with a 2020 copyright.

FileSizeNote
/core/rfb.js105,544 bytes, 2,988 linesStock noVNC core
/app/ui.js58,293 bytes, 1,715 linesStock noVNC interface
/core/websock.js10,593 bytesStock noVNC WebSocket layer
/vnc.html, /vnc_lite.html, /utils/novnc_proxyn/aClient pages and proxy helper

The listing shows static client files next to the proxy. It does not show a live VNC session or any cookie import. Replay from the same IP would also need a running browser, a VNC server, a WebSocket bridge and a cookie transfer. None of these was observed, and a noVNC client alone does not replay cookies. No login lure or victim session was found, so Host B stays on the watchlist and its address is withheld.

What stops this kind of attack

If an operator did run a browser on the proxy VPS, IP-consistency and impossible-travel rules would not fire. The main defence is at sign-in: origin-bound passkeys. A passkey is tied to the real site's origin, so it will not sign in through a reverse proxy or a remote viewer. A proxy can also strip DBSC headers, so DBSC alone is not enough at sign-in.

ControlWhen it actsScope and limits
W3C WebAuthn / FIDO2 passkeysAt sign-in (primary AiTM and browser-in-the-middle defence)Origin-bound challenge defeats reverse-proxy and remote-viewer phishing
W3C DBSCAfter sign-in, at cookie refreshWork-in-progress Editor's Draft; browser pilots only
Entra ID Token ProtectionAfter sign-in, token bindingNative apps on Windows, iOS/iPadOS and macOS, plus limited browser preview support

5. Ligolo-ng Tunnel Relays and the YoroTrooper Cluster

In short

Ligolo-ng gives an attacker a routed tunnel into an internal network through one encrypted connection. Recently active relays sit on mainstream cloud and VPS networks and favour the default port 11601 or HTTPS 443.

How Ligolo-ng works

Older pivoting through SOCKS proxies forces operators to wrap each tool in proxychains. That approach hooks sockets through LD_PRELOAD, handles ICMP and UDP poorly, and creates many short connections.

Ligolo-ng works differently. The relay on the operator side creates a virtual Layer-3 TUN interface (through ip tuntap, its own interface_create command, or wintun.dll on Windows). It handles the packets in Google's userland gVisor netstack.

The relay then tells an unprivileged Go agent on the target host to open ordinary TCP and UDP sockets. All traffic travels over a single TLS or WebSocket stream (MITRE ATT&CK T1572). Figure 4 summarises ports, hosting and the YoroTrooper overlap.

Figure 4: Active Ligolo-ng relays cluster on default port 11601 and HTTPS port 443
Figure 4: Active Ligolo-ng relays cluster on default port 11601 and HTTPS port 443

Ports

PortRelays (share)ProfileTypical hosting
TCP 11601129 (37.2%)Unchanged default relay portOVH SAS (IN, FR), DigitalOcean (SG, DE), Contabo (DE), Hangzhou Alibaba (CN)
TCP 44394 (27.1%)HTTPS port to pass strict firewalls, with Let's Encrypt or self-signed TLSMicrosoft (US), Amazon.com (US, FR, KR), DigitalOcean (SG), Google (BE)
TCP 8020 (5.8%)Plain HTTP or WebSocket upgrade, or ACME HTTP-01 challengeAmazon.com (CA), DigitalOcean (US), Contabo (FR), WorldStream (NL)
TCP 8443 and 808014 (4.0%) each, 28 (8.1%) combinedAlternate HTTPS and web-proxy ports, often paired on one relayMicrosoft (US), Oracle (CL), Yandex.Cloud (RU), Google (US)
TCP 2083, 11602 to 11607, 444519 (12 on 2083, 5 sequential, 2 on 4445)CDN-compatible TLS port, multi-agent ports, a custom portMicrosoft (2083), HOSTING INDUSTRY (11601 to 11607), Contabo (4445)

Countries and hosting

The United States (94), India (40), France (29), Germany (27), Russia (21) and Singapore (19) hold 230 of the 347 relays (66.3%). The relays sit inside mainstream cloud and European VPS networks, so their outbound TLS blends with ordinary SaaS traffic.

Hosting provider (ASN)Relays (share)Distinct IPsPattern
OVH SAS (AS16276)33 (9.5%)33Single-listener VPS relays in India, France, Singapore and Canada
Microsoft (AS8075)28 (8.1%)17Azure relays pairing 443, 8443 and 2083 on shared IPs
DigitalOcean (AS14061)26 (7.5%)25Single-port droplets on 11601 and 443 in US, SG, DE and GB
UpCloud (AS202053 / AS25697)22 (6.3%)2214 on UpCloud Ltd and 8 on UpCloud USA, mostly on 11601
Amazon.com (AS16509 / AS14618)21 (6.1%)21Single-listener EC2 relays, mostly 443 and 80
Contabo (AS51167)21 (6.1%)15Multi-port European VPS nodes on 11601, 443, 80 and 4445

None of the 347 relays is confirmed: 5 are suspected and 342 are candidates. At the port and banner level, a Ligolo-ng relay looks the same as an authorized red-team relay. Five of the hits sit inside a U.S. public-sector network and are held out, since they may be banner collisions, authorized testing or exposure on the victim side.

YoroTrooper overlap

Cisco Talos documents YoroTrooper espionage campaigns against organisations in the Commonwealth of Independent States (CIS), Turkey and Europe. The group uses spear-phishing, credential harvesting, Python information stealers and reverse shells.

In a 2026 assessment of the Kazakhstan-nexus group UAT-6275, Talos noted the two may sit in a broader CIS-targeting cluster. Talos still tracks UAT-6275 separately, because it saw no direct indicator or tool overlap.

Separately from Talos, third-party feeds tag 83 IPv4 addresses as YoroTrooper-related infrastructure (tags NIK C2, Chvost, Laplas, PickleShell and SSDPShell). Of 50 of these hosts that we reviewed, 11 carry feed labels for Ligolo-ng, dated 14 to 15 October 2026. They sit on DigitalOcean (6), MTS PJSC (3), Contabo (1) and Enterprise Cloud Ltd. (1).

Only 1 of the 11 appears among the recently active relays. The other 10 showed no recent Ligolo-ng banner. Feed tags are not attribution, so these addresses are not published.

6. Keeping Vendors and Dual-Use Tools off Blocklists

In short

Raw feeds label commercial proxies, training platforms and appliances as hostile. Separating them, and separating watchlists from blocklists, keeps enterprise controls accurate.

Figure 5: Layered triage holds out vendor collisions before passkey controls
Figure 5: Layered triage holds out vendor collisions before passkey controls

Seven rules shape every number and indicator in this report. Each one is also a rule a SOC can apply to its own feeds.

  1. Awareness-training certificates are held out. Four confirmed GoPhish rows on 2 AWS IPs carry a training vendor's certificate name. The 83 other confirmed GoPhish rows (84 in global records) stay on the watchlist. Unverified GoPhish banners (23,119 globally, 1,113 in the active set) are held.
  2. Commercial proxy and appliance certificates are held out. Of 156 confirmed Mitmproxy rows, 44 (37 DigitalOcean and 7 Constant/Vultr) carry commercial proxy certificate names and 2 carry CN=FortiGate. The other 112 (110 without FortiGate) stay on the watchlist. Unverified Mitmproxy rows (1,721 globally, 1,044 in the active set) are held.
  3. Authorized testing servers and public-sector hits are excluded. One authorized out-of-band testing server is left out of all hostile counts. The 5 Ligolo-ng hits in a U.S. public-sector network are excluded from adversary VPS claims and anonymised.
  4. Ligolo-ng listeners stay hunt leads. The 347 relays (342 candidates, 5 suspected) are never promoted to a blocklist without endpoint egress evidence.
  5. Stock open-source files are never blocked alone. The noVNC (rfb.js, ui.js) and Evilginx2 (phishlet.go, gophish.go) hashes only show co-hosting context. Blocking them alone would break legitimate Proxmox VE, OpenStack and VDI consoles.
  6. Red-team provenance is considered. Host A holds a keystroke receiver, an Evilginx2 store and Cobalt Strike files together. Its authorization is unknown and a red-team exercise cannot be excluded, so its address is withheld. Its custom script hashes stay published as tool indicators. Host B and Host C stay on watchlists with addresses withheld.
  7. Victim and third-party data are withheld. The 21 visitor IPs in the Evilginx2 blacklist, the consumer portal domains, the Cronicle server address, merchant paths, API keys, tested credentials and all non-confirmed listener addresses are withheld.

In total, 24,840 unverified dual-use records are kept off blocklists. Vendor certificate names should only lead to exclusion when the certificate chain is verified as trusted. A self-signed certificate that copies a vendor name is a reason to look closer, not to exclude.

7. What Defenders Should Do

Identity

  • Require origin-bound W3C WebAuthn passkeys with device-compliance Conditional Access as the primary control against AiTM proxies and noVNC lures.
  • Turn off legacy IMAP and SMTP basic authentication.
  • Add Entra ID Token Protection on supported native apps (Windows, iOS/iPadOS, macOS, plus limited browser preview support) and pilot browser DBSC on supported sites.

Network

  • Block outbound traffic to TCP 11601 to 11610.
  • Run the Sentinel query that links risky sign-ins to tunnel connections within 24 hours.
  • Test the eight Suricata signatures in a lab, then deploy them at the perimeter, in EASM tooling and on decrypted proxy taps. Exclude sanctioned VDI consoles from SIDs 9602604 and 9602608.
  • Alert on text/plain POSTs to a /k path whose body starts with a bracketed field label.

Endpoint

  • Test, then deploy, the five Sigma rules for Ligolo-ng agents on Linux and Windows, unsanctioned wintun.dll loads, and noVNC or Evilginx2 processes on cloud VMs.

Triage

  • Keep commercial proxy, awareness-training and FortiGate certificate rows, port-3333 pools and stock open-source hashes off blocklists.
  • Keep passive listeners and dual-use workspaces on watchlists until a lure or egress signal corroborates them.

8. Detection Rules

The full rules are in the Detection rules tab. No rule was run on a live engine; test each one in your own environment before deployment.

Rule fileWhat it catches
suricata_aitm_sendbeacon_and_ligolo.rulesSuricata 7: eight signatures for keystroke beacons, exposed Evilginx2 and noVNC, M365 enumeration bursts, Ligolo-ng ports, non-browser WebSocket upgrades and VNC greetings
sentinel_aitm_token_replay_and_ligolo_hunt.kqlMicrosoft Sentinel and Defender XDR: tunnel connections linked to risky Entra ID sign-ins by the same user
sigma_evilginx_novnc_and_ligolo_agent.ymlSigma (5 rules): Ligolo-ng agents, unsanctioned wintun.dll loads, noVNC and Evilginx2 or cloner processes
yara_aitm_keylogger_and_phish_toolkits.yarYARA (4 rules): the keystroke script, the receiver, the cloner, and the M365 and credential-reuse scripts

How the Suricata signatures work

SIDLooks forWhere to deploy
9602601Outbound text/plain POST to a path ending in /k, whose body starts with a short bracketed field labelPerimeter or decrypted web proxy
9602602Request URIs for exposed Evilginx2 runtime, phishlet, core or database foldersEASM and hunting
9602603Inbound script that contains the Beacon API name, a deletion token and a page-unload event name close togetherEASM and decrypted proxy
9602604Inbound HTTP 200 response with noVNC client pages or core/rfb.jsEASM and hunting; exclude sanctioned VDI
9602605Four or more GetCredentialType POSTs from one source in 120 seconds, from a client without the browser sec-fetch-site headerSSL-inspecting proxy or cloud tap
9602606Outbound TLS client hello to ports 11601 to 11610Perimeter; a port-policy alert
9602607Outbound WebSocket upgrade without Origin or Cookie headers, rate-limited per sourcePerimeter; a hunting rule
9602608Inbound raw VNC RFB 003.00x server greetingEASM and hunting; exclude sanctioned VDI

SID 9602605 tolerates both compact JSON and Python json.dumps() spacing in the request body. It excludes browsers by the missing header, not by user agent.

How the Sentinel query works

The query looks back 14 days. It first builds a list of successful Entra ID sign-ins that are risky: medium or high risk, risk events such as anomalousToken, or a sign-in from a high-concentration relay network on a non-compliant device. The relay network list holds six ASNs, including OVH (16276), Contabo (51167) and UpCloud (202053 and 25697).

It then takes device connections to ports 11601 to 11610, or to 443, 8443, 2083, 8080 and 4445 from non-browser processes whose command line names Ligolo-ng flags or wintun. It joins the two on the user account and keeps a sign-in only if the tunnel started within 24 hours of it.

The output labels each row. "Identity-correlated tunnel" means a risky sign-in matched. "Port policy only" means a connection to a default Ligolo-ng port with no matching sign-in. Connection counts are deduplicated so the join does not inflate them.

How the Sigma rules work

RuleFires when
1. Ligolo-ng Linux agentA Linux process starts with a connect argument (host:port, https:// or wss://) under a default agent name, or a renamed binary carries Ligolo flags such as -ignore-cert, -accept-fingerprint, -retry or a :1160x port
2. Ligolo-ng Windows agentThe same logic on Windows; curl --connect-timeout and openssl s_client are excluded
3. Wintun DLL loadwintun.dll loads into a process that is not an approved VPN client executable (exact paths for WireGuard, Tailscale and Cloudflare WARP; add your own)
4. noVNC on a cloud VMnoVNC or websockify starts on cloud or VPS hosts
5. Evilginx2 or cloner on a cloud VMEvilginx2 or a switch_site.sh cloner starts on cloud or VPS hosts

A renamed agent that dials plain host:443 without Ligolo flags will not trip rules 1 and 2. The Sentinel query covers that case through network correlation.

How the YARA rules work

Each YARA rule is limited to files under 64 KB and needs all of its strings to match.

YARA ruleTargets
ODIN_AiTM_SendBeacon_PreSubmit_Keylogger_JSThe 811-byte keystroke script
ODIN_AiTM_CORS_Keystroke_Receiver_PythonThe Python clone server and /k receiver
ODIN_AiTM_OneClick_Login_Site_Cloner_ShellThe Bash portal cloner
ODIN_M365_GetCredentialType_And_Lateral_Auth_Toolkito365_test.py or lateral_auth.py

How to sort a host

The triage logic never blocks on a file path alone. It needs reviewed file content, or live interception or egress evidence. It checks for exposed hostile files before it applies any vendor exclusion, so a self-signed certificate that copies a vendor name does not get a free pass.

What the host showsAction
Vendor certificate name with a verified trusted chain, FortiGate appliance, authorized testing serverExclude
Vendor certificate name with unknown or failed chain trustWatchlist
Unverified GoPhish on 3333, unverified Mitmproxy, Ligolo-ng on 11601 aloneHold or hunt
A hit inside a public-sector networkReview only
Confirmed passive listener, co-hosted proxy and noVNC, dual-use recon workspace, workbench paths without reviewed contentWatchlist
Ligolo-ng relay with endpoint egress, keystroke receiver with reviewed content, active browser-in-the-middleBlock

Vendor names in this logic are placeholders. Supply your own allowlists, and mark a chain as trusted only when you have checked it.

9. Indicators

The Indicators tab holds the 3 custom script hashes from Host A, whose address is withheld. The other 25 records in the selected set are watch, hunt or context entries described below.

TypeCountAction
Host A (address withheld)1Hunt by its file hashes
Custom script hashes (SHA-256) from that host3Hunt
Watchlisted hosts (Host B, Host C)2Watch; addresses withheld
Host C script and lure hashes (SHA-256)4Watch
Framework-confirmed Mitmproxy and GoPhish listener profiles9Watch; addresses withheld
Stock open-source hashes (Evilginx2, noVNC)4Context only; do not block alone
Candidate Ligolo-ng relay profiles5Hunt egress only; addresses withheld

Hunt

IndicatorPortWhat it is
dc89160b8501065099a69bc6a8f34ce085774878387d071f4e249eebb57a757bn/ak.js keystroke script
629fd5033240e77684003b632c3b247671ac2043a4e87ae7f420feb836551a45n/areceiver.py clone server and receiver
7da269d7b5b55d409b71702dc53d4329bbeb0afb0540ab3d98896c3625836c54n/aswitch_site.sh portal cloner

Context only (stock open-source files)

SHA-256FileWhere found
00d134e5be68c83a13c35101cd26aca6efe6af608144c35add5edf3a4139f336Evilginx2 core/phishlet.go (29,710 bytes)Host A
6097864f26657ac5add013af5e17f508678dec3eb57548dbf37ea09e2a02a78cEvilginx2 core/gophish.go (2,988 bytes)Host A
331109966c80bef620bcd09242d43dbc05fa83cb44eba439ad20148f86aa479enoVNC core/rfb.js (105,544 bytes)Host B
70ba76ca86ceaea27dc338bbfee4cebcfef1160812c8d410241f3667dff4864cnoVNC app/ui.js (58,293 bytes)Host B

Three of the four Host C hashes are in the Host C table in Chapter 4. The fourth, for the cloned login page /mp_login.html, is withheld because the page copies a third-party site. The nine watchlisted listener profiles sit on hosting providers in the United States, Vietnam, Germany, Russia and Indonesia. The five candidate relays sit on providers in the United States, China and India.

Withheld: the addresses of Host B, Host C, the nine watchlisted listeners and the five candidate relays; the 83 YoroTrooper-tagged addresses; the 21 visitor IPs; and all third-party portal, merchant and certificate names.

MITRE ATT&CK mapping

TacticTechniqueWhere it appearsDetection
Credential Access, CollectionT1557 Adversary-in-the-MiddleEvilginx2 tree on Host A; confirmed Mitmproxy listenersSID 9602602; Sigma rule 5
Credential AccessT1539 Steal Web Session CookieEvilginx2 session store; Host B proxy and noVNC pairingSIDs 9602604, 9602608; Sigma rule 4
Command and ControlT1572 Protocol Tunneling347 recently active Ligolo-ng relaysSIDs 9602606, 9602607; Sentinel query; Sigma rules 1 to 3

10. Scope and Limits

  • Window. Findings come from a two-week observation window ending 18 November 2026, with a data cutoff of 16 November; most listeners were last seen from 12 November onward. The workbenches were observed on 13 and 16 November 2026.
  • Global records are totals, not a timeline. The 26,329 global banner records count every record on file for each framework. They do not mean Odin watched these hosts for a long period, and they support no trend claims.
  • Confirmed is not hostile. A framework-confirmed listener runs the software. It may be a red team, a researcher or a training platform. That is why confirmed listeners sit on watchlists.
  • Files show intent, not outcomes. The scripts on all three workbenches show what they were written to do. No target-side result was verified, and the contents of the 136,742-byte log file are not established.
  • No campaign link. Nothing links Host A, Host B and Host C to each other. They are three separate findings.
  • Feed tags are not attribution. The YoroTrooper overlap rests on third-party feed labels. It is not tied to the Talos reports by a matching indicator.
  • Rule testing. Every published rule file loads without errors in its engine (Suricata 7.0.12, YARA-X, pySigma and Microsoft's Kusto.Language analyzer, checked on 20 November 2026). The rules were not replayed against live traffic, so test and tune them in your own environment before production use.

11. Sources

  1. MITRE ATT&CK T1557: Adversary-in-the-Middle. https://attack.mitre.org/techniques/T1557/ (verified 18 November 2026)
  2. MITRE ATT&CK T1539: Steal Web Session Cookie. https://attack.mitre.org/techniques/T1539/ (verified 18 November 2026)
  3. W3C. Beacon specification. https://www.w3.org/TR/beacon/ (verified 18 November 2026)
  4. W3C. Web Authentication Level 3. https://www.w3.org/TR/webauthn-3/ (verified 18 November 2026)
  5. W3C. Device Bound Session Credentials, Editor's Draft. https://w3c.github.io/webappsec-dbsc/ (verified 18 November 2026)
  6. Microsoft. Token Protection in Entra ID Conditional Access. https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-token-protection (verified 18 November 2026)
  7. MITRE ATT&CK T1572: Protocol Tunneling. https://attack.mitre.org/techniques/T1572/ (verified 18 November 2026)
  8. nicocha30. Ligolo-ng repository. https://github.com/nicocha30/ligolo-ng (verified 18 November 2026)
  9. Cisco Talos. YoroTrooper espionage campaign targets CIS, Turkey and Europe. https://blog.talosintelligence.com/yorotrooper-espionage-campaign-cis-turkey-europe/ (verified 18 November 2026)
  10. Cisco Talos. Kazakhstan-nexus UAT-6275 targets CIS government organizations. https://blog.talosintelligence.com/break-the-silence-kazakhstan-nexus-uat-6275-targets-cis-government-organizations/ (verified 18 November 2026)

What is…?

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

All explanations

Frequently asked questions

How does a phishing kit capture a password nobody submitted?

A normal phishing clone only receives a password when the user clicks Submit. If the user types half a password, notices the wrong web address and closes the tab, the attacker gets nothing.

What do the other two workbenches show?

The other two servers are on our watchlist, and their addresses are withheld. Why does that pairing matter? When an attacker replays a stolen cookie from a laptop, identity systems often flag the new location. Replaying it from a browser on the proxy server itself would keep the same address. But that also needs a running browser, a VNC server and a connecting bridge. None of these was observed on Host B.

Where do the tunnel relays run?

Once inside, attackers need a route into internal networks. Ligolo-ng provides one. Its relay creates a virtual network interface on the attacker's server. A small agent on the victim host then opens ordinary connections, all carried over a single encrypted stream.

Why not just block every listener?

Because many of them are not hostile. Among confirmed Mitmproxy listeners, 44 carry the certificate name of a commercial proxy network. Four confirmed GoPhish listeners carry the name of a security-awareness training vendor. Two more carry FortiGate appliance certificates.

Defender kit

Indicators of compromise

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

IDTypeIndicatorPortContextAction
R06-IOC-01 SHA-256 dc89160b8501065099a69bc6a8f34ce085774878387d071f4e249eebb57a757b 8000 Pre-Submit DOM Keystroke sendBeacon Streamer SHA-256 Hunt
R06-IOC-02 SHA-256 629fd5033240e77684003b632c3b247671ac2043a4e87ae7f420feb836551a45 8000 CORS-Wildcard Keystroke & Redirect Receiver SHA-256 Hunt
R06-IOC-03 SHA-256 7da269d7b5b55d409b71702dc53d4329bbeb0afb0540ab3d98896c3625836c54 8000 One-Click Login Clone & Asset Rewriter Script SHA-256 Hunt

25 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.

Sentinel KQL sentinel_aitm_token_replay_and_ligolo_hunt.kql 44 lines
Download
Sigma sigma_evilginx_novnc_and_ligolo_agent.yml 181 lines
Download
Suricata suricata_aitm_sendbeacon_and_ligolo.rules 33 lines
Download
YARA yara_aitm_keylogger_and_phish_toolkits.yar 73 lines
Download