New Intent Workflows and AI Instruction Intercept: a named person on every critical request, and no hidden instructions for your AI assistant.New: Intent Workflows and AI Instruction Intercept. See what’s newSee

Threat research

What the Scanner Sees vs. What You See: Network-Based Cloaking in an Email Phishing Campaign

The same phishing link showed a decoy to a cloud scanner and redirected to a tracking service in residential and mobile tests.

Featured image illustrating network-cloaked phishing, showing how the same malicious email link presents an access-blocked decoy to a security scanner while redirecting another visitor through a different path.

In late October we reviewed a set of verified phishing-style messages that looked, at first glance, like routine "your cloud storage is full" and "claim your prize" spam. They were not routine in construction. Each message shows the reader one image and one link. Behind that image sits a hidden block of borrowed email threads and random tokens. The link leads to a randomly named host whose behavior depends on who is asking: a cloud-hosted scanner gets a decoy page, while a residential or mobile visitor is passed on toward an affiliate tracking network.

This Guide walks through the campaign layer by layer, from what the recipient sees to the infrastructure behind it.

Infographic showing observed phishing kit activity, including 7 registered domains, 6 distinct pasted email threads, 6 origin IPs, approximately 50 Microsoft 365 relay tenants, and 52 sender domains, with activity recorded from October 9 to November 1.
Phishing Kit Infrastructure and Activity Overview: A breakdown of the observed attack infrastructure, including registered domains, email threads, origin IP addresses, Microsoft 365 relay tenants, and sender domains, highlighting the scale and distribution of the campaign between October 9 and November 1.

Five findings:

  1. Message layer. Every message is HTML-only. The reader sees one image inside one link. Hidden blocks carry the subject line, rule lines, random tokens and a pasted real email thread.
  2. Link layer. Links use per-message random hostnames with an unusual internationalized label that is built the same way every time, leaving a constant fingerprint.
  3. Network layer. Datacenter visitors are sent to a decoy. Residential and mobile visitors move on, sometimes through a JavaScript gate, to a tracker that answered our analyst with an empty response.
  4. Sender layer. Mail arrives through about 50 Microsoft 365 tenant IDs but from only six origin IPs.
  5. Content provenance. The pasted threads look harvested from real mailboxes rather than leaked publicly.

What the recipient sees: one image, one link

Every message we examined is HTML-only: no plain-text part, no attachment. The visible body is a single image wrapped in a single link. The image is delivered in three ways: embedded inline in most messages, pulled from a remote image host through a non-standard <image> tag in a large minority, or set as a CSS background in a few. The campaign draws on a small pool of image files, about two dozen in total.

There are two lure families. The first imitates a cloud-storage provider's "storage full, update your payment" notice, in English and French. The second is a set of consumer-scam funnels: a fake news-site health advert, a patient-portal holiday notice, prize and survey offers, a parking-app notice. Sender display names are brand-like and sit on unrelated domains.

Collage of phishing and scam examples showing fake Google Gmail and Microsoft reward notifications, a fraudulent survey offering a luxury watch, and a misleading EMSense promotional advertisement with before-and-after imagery.
Examples of Deceptive Online Offers and Phishing Lures: Suspicious reward notifications, fake brand promotions, and misleading advertisements use familiar branding, attractive incentives, and persuasive messaging to encourage users to click links or share information.

The cloud lures show how little effort went into them: the "account number" printed in the image is the same in every message that uses a given template, where a real notice would be personalized.

Lure theme also tracks sender infrastructure. The cloud-storage lures come mostly from education-sector domains, while the health and retail lures come mostly from random-subdomain hosts. We read that as two sending pools feeding two lure templates, and no more than that.

Many messages also carry a Sender: header with a random name on a domain that, as far as we can tell, is not registered, so mail clients that surface "on behalf of" wording show an odd line above the message.

Screenshot of a suspicious cloud storage phishing email with the subject “Your Cloud Storage Is Full — Act Now or Risk Losing Files,” showing an unfamiliar sender domain, an external email warning, and an urgent message designed to create fear of data loss.
Cloud Storage Phishing Example: A suspicious email impersonates a cloud storage notification and uses urgency and fear of losing files to encourage action. The unfamiliar sender domain and generic cloud-update identity are potential phishing warning signs.

What a text scanner sees: the hidden block

Open the HTML and the picture changes. Alongside the image-and-link there is a container hidden with visibility:collapse or display:none in nearly every message. Inside it are the subject line repeated, runs of rule characters, a pasted real email thread (one of six distinct threads reused across the set) and, in a good share of messages, random pipe-delimited token strings mixed with @_# character runs.

Some messages have no visible text at all. In a number of them the kit's anchor tag sits before the <body> element, a placement that some parsers handle differently from a link inside the body.

comparing what a recipient sees in an image-based phishing email versus what a text parser detects, including hidden HTML containers, repeated subject lines, six pasted email threads, and random pipe-delimited tokens.
A recipient sees a simple image-based phishing lure, while a text parser may encounter hidden HTML content, repeated subject lines, and unrelated email threads. This technique illustrates how attackers exploit differences between email rendering and text extraction to evade traditional phishing detection.

Two further regularities help with grouping. The subject carries a token of the form ID:###-XX-### (three digits, two letters, three digits) in nearly every message, and in many a different token of the same shape also appears in the body. A leftover empty image part named like a pasted Outlook signature appears in a large share of messages, which suggests the threads were pasted from a mail client rather than generated.

A note on rendering. It is tempting to say the thread is "invisible to the reader". Our analyst's mail client drew the collapsed thread below the lure image anyway, after a horizontal rule and a "Forwarded message" label.

The accurate statement is that the content is marked hidden with CSS, and whether the reader sees it depends on the renderer. The design is consistent with an attempt to dilute or disguise the text a parser extracts.

The kit link: a fingerprint in the hostname, a parameter string in the path

The link in nearly every message uses an internationalized hostname label. At first glance that suggests a homograph attack, but it is not one. The label is a random string of 8–10 letters with three Latin-1 characters inserted. It imitates no brand and mixes no scripts. What it does produce is a punycode label that always ends in the same suffix, -95a2o93b, in every such link. The suffix is a fingerprint of how the generator builds labels.

The kit uses a different random hostname in almost every message, spread over seven registered domains: paraiadauehungdai[.]com, greshika[.]com[.]np, nagawinwin[.]com, gragongold[.]com, banjirpt2[.]com, and two others that contain a software product name. We checked registration data on 2 Nov. Four were registered in 2025–2026; the two product-named domains date from April 2023 and sit on Cloudflare name servers. We saw nothing tying a vendor to the campaign.

The path is more interesting than the host. Click links start with a 4 and beacon images with a 3, followed by dot-separated integers. To read them we wrote a small decoder script. It splits each numeric path into its parts and compares them with the sub-ID parameters in the tracker URLs we captured. The first seven numbers carry three identifiers.

The last of them is sub1, a source ID fixed for each kit domain family (for example 10 on one family and 13 on another). The second number and the first, joined with a hyphen, form sub2. The next three numbers, joined, form sub3. Everything after that is random padding. On every numeric path we tested, the decoder reproduced the tracker's sub-IDs exactly.

Technical diagram showing how a phishing kit encodes tracking parameters in URL paths, using three sub-ID fields to identify domain families, distinguish tracking values, and correlate phishing activity across different registered domains.
Analysis of a numeric URL path reveals three sub-ID fields used to organize campaign activity. Shared identifiers across different domains help researchers link related phishing messages and infrastructure, even when the hostnames differ. Similar URL-pattern analysis is used in phishing campaign investigations.

The decoder showed three things. The set contains dozens of distinct sub2 values and many more distinct sub3 triples. A message's click link and its beacon image carry the same seven numbers, so every link in a message points to one set of affiliate IDs.

And the IDs link messages that the hostnames hide: the first two messages below sit on the same registered domain and the third on a different one, yet all three share the sub3 prefix 283-76894.

comparing three phishing kit messages across two domains, showing their seven-number URL paths and decoded sub1, sub2, and sub3 tracking parameters used to identify related phishing activity.
A comparison of three phishing messages reveals how numeric URL parameters encode tracking identifiers. The shared sub3 prefix (283-76894) helps link messages across different kit domains, while variations in sub2 and sub3 distinguish individual tracking values.

The structure of the link, and what the tracker received, points to a traffic-selling funnel rather than a credential-theft page.

What a network scanner sees: cloaking by visitor IP

The same link gives different answers depending on the network it is fetched from. This is the part we find most useful for defenders.

Layer 1: a network check at the kit server. In automated detonations from a cloud provider's network, the kit consistently returned the same decoy page, with text announcing that the visitor's network was blocked. We repeated the fetch by hand from residential and mobile connections. The server did not return the decoy; it redirected onward. Those manual runs exited through Cloudflare edge locations. The simplest reading is a check of the visitor's network against a list of datacenter ranges. The server also echoed the visitor's IP in a response header, and its headers indicate a limit of 500 requests an hour per IP. The responses are consistent with a Go web service built on the Gin framework.

Layer 2: a JavaScript bot gate on some hosts. On at least one host the first response is a 200 page that does nothing visible. Its script requires cookies to be enabled and a user agent that does not match a hard-coded list of crawlers. It then waits one second, hops to a /t/ path, waits one second more, and redirects. A scanner that does not execute JavaScript stops at "redirecting...".

Layer 3: the tracker declines. The redirect lands on hosts that behave like Everflow-style tracking endpoints, identifiable from their request-ID header and URL shape. Captured exchanges show an HTTP 204 with no body returned to our analyst. The empty response is consistent with the tracker filtering traffic that does not match an offer's targeting, such as location or device.

Flowchart showing how a phishing kit delivers different responses based on visitor network type, including a decoy page for datacenter IPs and JavaScript bot checks, affiliate tracking, and an HTTP 204 response observed during residential or mobile testing.
Same Link, Three Different Answers: This phishing kit uses network-based filtering and bot checks to distinguish visitors. Datacenter requests receive a decoy page, while the residential or mobile testing path proceeds through additional checks and tracking, illustrating how attackers can conceal malicious activity from automated security analysis. Similar traffic-filtering techniques have been documented in threat research.
Infographic showing three phishing kit traffic-filtering layers—network checks, JavaScript bot detection, and affiliate tracking—alongside observed behavior and evidence, including datacenter decoy pages, conditional redirects, and HTTP 204 responses.
Phishing Kit Traffic Filtering: Behavior and Evidence. The analysis reveals how network-based filtering, JavaScript checks, and tracker responses influence phishing link behavior. These techniques can conceal malicious destinations from automated security scanners while allowing different responses for residential and mobile visitors.

Senders: many tenants, few origins

Mail passes through about 50 Microsoft 365 tenant IDs, but the raw headers trace to only six origin IPs: 213[.]246[.]56[.]37 (hosting in France, the largest source), 91[.]235[.]234[.]140 (Estonia), 103[.]4[.]219[.]32 (Thailand), 185[.]246[.]84[.]73, 103[.]91[.]191[.]184 and 185[.]39[.]18[.]76.

The pivot is that two of those IPs, 103[.]4[.]219[.]32 and 103[.]91[.]191[.]184, also resolve the kit's landing subdomains through wildcard DNS. The hosts that sent the mail also serve the landing pages, and the same small set of VPS hosts feeds all of the roughly 50 relay tenants.

phishing campaign infrastructure connecting six origin IP addresses to approximately 50 Microsoft 365 tenant IDs and 52 sender domains, with reused MIME boundaries, email header patterns, and shared phishing kit hosting infrastructure.
Few Origins, Many Tenants, Rotating Names: StrongestLayer's threat research identifies six origin IPs linked to approximately 50 Microsoft 365 tenant IDs and 52 sender domains. Shared MIME boundaries, recurring email header patterns, and overlapping phishing kit hosting provide evidence connecting the campaign's distributed infrastructure.

The headers carry their own fingerprints: three MIME boundary strings reused across nearly all messages; To equal to Cc in every message; no X-Mailer; the random Sender: names noted above; and two Message-ID shapes, one at a webmail host and one at a single fixed Outlook host. There are 52 From domains, and the From and Return-Path domains share a registered domain in every message. DKIM d= equals the From domain in most of them.

The From domains are mostly aged, and several are education-sector domains rather than look-alikes of any brand. Rotating sender names and domains make a poor pivot here; the origin IPs, the boundary strings and the header shape do better.

Where the pasted threads came from

The hidden threads are real, not synthetic. There are six, each pasted unchanged into many messages: a school fee reminder, a business plan, a parent's complaint, a forwarded tax notice, a family email and a conference forward. They come from at least three mailboxes. Some contain personal details, which is why we do not quote them.

Three observations favor harvesting from live mailboxes over a public leak. The tax-notice forward was already in attack mail five days after the date on it. Three of the threads trace back to a single school mailbox. And a search of our retained detections shows threads joining the rotation in batches. We found no public copy of any thread and no report of a breach at the organizations involved.

We read this as harvesting from live mailboxes. We did not research any individual, and we deliberately describe the threads without naming the people or organizations behind them.

Timeline: how long it ran

The messages we reviewed arrived between 24 and 29 October. A wider look at related detections shows the same construction from 9 October to 1 November, across multiple organizations.

Timeline infographic showing an observed phishing campaign from October 9 to November 1, 2026 (UTC), with the main activity occurring October 19–29 and peak activity recorded on October 22 and 25.
Phishing Campaign Activity Timeline (October–November 2026): StrongestLayer's threat research traces the same phishing message construction from October 9 through November 1, with concentrated activity between October 19 and 29. October 22 and 25 were the busiest observed days, highlighting the campaign's most active period.

Messages with the same construction appear over at least three weeks, in bursts. The busiest days in our data are 22 and 25 October.

How TRACE detected it

TRACE, our reasoning-based email security engine, detected and classified every message in this campaign as phishing. It does not match one indicator. It collects the artifacts of a message, reasons over them together, and writes out why it reached its verdict. The campaign is a good test of that approach, because each individual artifact looks weak or even harmless: a single image, a link to an unfamiliar host, a school sender, a pasted business thread. The case rested on how the artifacts fit together.

seven email threat indicators examined by StrongestLayer's TRACE engine, including sender identity, claimed identity, message content, embedded images, redirect paths, link behavior, and language and intent in a phishing campaign.
StrongestLayer's TRACE engine evaluated sender history, identity mismatches, hidden message content, suspicious URLs, redirect behavior, and deceptive language to identify signals of phishing. The analysis demonstrates how contextual and intent-based reasoning can reveal threats beyond conventional email security indicators.

Three things in the campaign show why combining these matters.

The decoy became evidence. The link check returned a malicious verdict for most messages. Our tests explain why: the kit shows a "your network is blocked" page to cloud-hosted visitors, and TRACE read that text as a sign of automated attacker tooling. So the network cloaking that hides the real destination from a scanner also left a marker on the scanner. We do not want to oversell this. It worked because the check came from a cloud network, and a visitor on a home connection would have been passed on.

Some messages did not need the link. For a handful of messages the link check was inconclusive, because it timed out or stopped at the cloaking. TRACE still reached the same verdict from the rest. It read the lure from the embedded image, saw that the sender had no connection to the provider the image imitates, and saw that the pasted business thread had nothing to do with a storage notice. In one case it described the message as a split-surface design: harmless text on the surface, the real lure in the image. This is the case where reasoning over several artifacts matters, because no single one of them was conclusive.

The hidden thread was a clue, not a cover. The pasted threads are meant to make a message look like genuine correspondence. TRACE treated the mismatch between the thread and the notice around it as a reason for suspicion, not as a reason for trust.

What defenders can take from this

  • Inspect the rendered message and compare it with the extracted text. A message whose visible content is one image while its source holds paragraphs of unrelated text is worth a flag of its own.
  • Look for hidden collapsed containers that carry a copy of the subject, rule lines and token strings.
  • Fingerprint on structure rather than content: the constant punycode suffix, boundary-string reuse, To equal to Cc, and a single image-wrapped link.
  • Detonate from more than one network type. Cloud-only detonation was shown a decoy here. A clean result from one vantage point is not a clean result.
  • Pivot on origin IPs and the VPS hosts behind the relay, not on rotating sender names.
  • Treat a link-protection timeout or a 204 as inconclusive, not as benign.

Scope of this analysis

This Guide covers verified phishing detections received between 24 and 29 October, with a wider look at related detections from 9 October to 1 November. The findings on network behavior come from automated detonations from a cloud network and manual runs from residential and mobile connections.

Final Thoughts

The page that stopped the cloud scan became part of the evidence. When a link check ended early, the request in the image, the sender’s relationship to the claimed service, and the surrounding correspondence helped explain why the email was suspicious. Record the last response the scanner observed and the evidence that supports the email verdict. That gives the next analyst a clear account of what the investigation established and what remains unresolved.

FAQ

What is network-based cloaking in phishing?

It’s when a site changes the page or redirect it serves based on the network a request comes from. The results in this case were consistent with network-based filtering. The exact routing rule remains unverified.

What did HTTP 204 mean in this investigation?

The tracking endpoint returned a response with no body. That was where the captured redirect chain ended. Any page a different visitor might have reached beyond that point remains unverified.

How did the team connect emails across rotating sender domains?

The team found shared originating IPs and recurring mail-header patterns. It also matched fields decoded from URL paths to tracking identifiers captured after the redirect, which helped group links across changing hostnames.

How did TRACE classify messages when link checks were inconclusive?

For the reviewed messages, TRACE examined the request in the image, the sender’s relationship to the claimed service, and whether the copied correspondence fit the request. Those signals supported a phishing verdict while the link path remained unresolved.

    ← All posts
    Free AI gap check

    Do you have an AI gap? Find out what’s getting through.

    See which generations of defense you’re running and what it takes to close the gap. Two minutes. No signup. Nothing changes in your environment.