What happened
On October 21, 2026, Huntress published research describing a malvertising campaign that began with a sponsored Google search result for the phrase "amazon business." According to Huntress, the paid result sent searchers to a fake Amazon sign-in page, then to a fake CAPTCHA that persuaded them to paste a command into Windows. That command started a multi-stage infection that ended with an information stealer and a pirated remote access tool on the computer. Huntress says the method struck at least three of its customers.
The researchers, Andrew Brandt and Sarah Reddish, report that Google has removed the malicious ad. They also report that the attacker infrastructure behind the later stages remained online and functional when they published. Dark Reading's October 21 report carried the same core findings and added interview comments from both researchers. It reports on Huntress's investigation rather than providing a separate measurement of the campaign.
Huntress does not report any compromise of Amazon's own systems. The campaign borrowed the Amazon Business name and the look of an Amazon sign-in page to make an ordinary procurement search feel routine. Huntress does not name the affected customers or give the dates of the three incidents. It also does not say how long the ad ran, how many people saw or clicked it, or where it was shown.
The path from search to command
Huntress describes the sponsored result as the top paid placement for "amazon business" at the time of the attacks. Its link pointed to a subdomain of an attacker-controlled domain that has nothing to do with Amazon, crafted so the start of the hostname resembled an Amazon orders address. From there, the visitor was redirected to a counterfeit Amazon sign-in page hosted on Google Sites. This article does not reproduce those addresses. Huntress lists them, defanged, in its indicator tables.
The fake page asked for an email address or phone number, the first step of a familiar Amazon login. After the visitor submitted it, a fake CAPTCHA appeared. Clicking the "I'm not a robot" button silently copied a command to the clipboard, and a pop-up told the visitor to paste it into the Windows Run box. This is the technique known as ClickFix. Visitors whose browsers identified as macOS received alternate Terminal instructions, although that payload address was offline when Huntress examined it.
The command was disguised. Huntress says its characters were broken up with commas, carets, hyphens and underscores that the Windows command interpreter ignores. Trailing spaces pushed the real command out of view so the Run dialog appeared to say only "Push the Enter key." Huntress does not say whether the email addresses or phone numbers typed into the fake form were stored or reused.
What the endpoint chain installed
The pasted command called finger, a networking utility older than the web that still ships with Windows. It fetched a script hidden in a remote ".plan" file, the kind of text profile finger was designed to display. Huntress says 68 blank lines pushed the script below the visible screen. That script downloaded a legitimate IronPython release from GitHub, saved it in the Temp folder with a .pdf extension, unpacked it and used it to run an encoded Python script.
Several layers followed. Huntress traced the encoded script to a loader, which retrieved a payload that security vendors identify as Doina, a modular malware family. Doina then pulled down a portable Node.js package with a heavily obfuscated script. That script decrypted and injected what Huntress describes as an alleged Lumma Stealer sample. The stealer looked for browsers, cryptocurrency wallets, password managers and email clients. A pirated copy of the NetSupport Client remote management tool was installed in a nonstandard folder. Its license and copyright details indicate it is at least eight years old.
The infrastructure findings matter beyond a single endpoint. Most command and payload domains had barbecue-themed names, and Huntress linked them through common hosting, DNS servers and shared payloads. The NetSupport client communicated with an IP address that otherwise hosts the website of a Detroit restaurant, a site Huntress says has not itself been observed in abuse. Each payload download returned an archive with a unique hash, and the payloads inside changed once or twice a day.
What is confirmed and what is not
The sponsored ad, redirect, Google Sites page, fake CAPTCHA, infection chain, three affected customers, ad removal and surviving infrastructure all come from Huntress's own investigation. Two labels carry qualifiers in the source. The Doina identification rests on what Huntress calls security vendor consensus, and Huntress calls the stealer "allegedly" Lumma. Huntress names no operator and makes no attribution. It does not say whether the advertiser account was created for the campaign or taken over from someone else.
Huntress's August analysis of a separate case found that a sponsored result for installing Claude on a Mac led to a shared conversation page on the real claude.ai domain and then to the MacSync stealer. That earlier campaign shows the same delivery idea with different infrastructure. Huntress does not connect the two operations, and this article does not either.
Where Outtake could have helped
This section is a hypothetical application of Outtake's published capabilities. It does not imply that Amazon, Huntress's customers or any affected organization uses or lacks Outtake, and it does not claim the infections would have been prevented. The relevant opportunity sits outside the endpoint: the ad, the lookalike destination, the hosted sign-in page and the infrastructure behind them. Outtake does not describe itself as endpoint protection, and it would not have removed malware from an infected computer.
Two kinds of organizations have a stake. Brand owners face attackers who buy their names as search keywords and copy their sign-in pages. Enterprises face employees who search for business tools during an ordinary workday. The analysis below mainly follows the brand-owner view, with a note on employee reports.
Spot the paid placement as brand abuse
Outtake Brand Protection lists paid advertising networks and domains among the channels it monitors. It describes detection across text and images, including fraudulent ads and phishing sites that clone a brand's visual identity. For a brand in Amazon's position, the useful signal is a sponsored result for the brand's own name whose destination is not an authorized brand domain or partner. Comparing the ad text, destination host and landing page design with official properties could support classifying the result as impersonation rather than ordinary competition.
Earlier identification matters here because the ad sat at the top of results for a routine work task. Every hour it stayed live gave more searchers a chance to click it. Huntress does not report when the ad first appeared, so there is no basis for saying how much earlier any monitor would have found it. Visibility would also depend on the keyword, location, device and timing settings the advertiser chose, and on whether the destination showed scanners a harmless page.
Connect the lure to the infrastructure behind it
Outtake's Brand Protection page describes tracing a flagged domain through shared hosting, registration patterns, analytics IDs and linked accounts. Its Recon Agent is described as expanding one signal across shared infrastructure to related assets. Applied to this campaign, an investigation could start from the ad destination, follow the redirect to the Google Sites page, then examine the domains the fake CAPTCHA tells a computer to contact.
Huntress built this kind of map itself, linking the barbecue-themed domains through hosting, DNS and payload relationships. A comparable map could support reporting the connected assets together instead of the ad alone. That matters because Google's removal of the ad left the later stages running. It also matters because payloads rotated daily, so the domains were more stable evidence than any single file.
Relationship evidence still needs care. Shared hosting alone does not prove common control. The Detroit restaurant IP shows how a legitimate business can sit inside an attacker's path. Each link in a map should carry its own evidence and observation date before it becomes the basis of a report.
Route evidence to the providers that can act
Outtake's platform describes evidence packaging, automated reporting, platform escalation and verified takedown, and the Brand Protection page names ad networks, hosts and registrars among the providers it works through. Applied here, that could mean three parallel reports. The first would go to the ad platform under Google's misrepresentation policy, citing the search query, ad text, destination and the click identifiers Huntress highlights. The second would go to the host of the counterfeit sign-in page. The third would go to registrars and hosts of the redirect and payload domains, with the evidence that ties each one to the chain.
Huntress makes a related point. Reporting ad-click identifiers to Google, it writes, can cut attackers off from the ad system, "at least, until they create a new account and start again." That caveat applies to any provider action. Removal decisions belong to the providers, and an operator can rebuild.
Link employee reports to the same campaign
Outtake Intake is described as taking in complaints from customers and employees, extracting indicators and connecting isolated reports into one campaign. In principle, an employee who reports that an Amazon login page asked them to paste a command could become a lead that links to other reports and to the external assets above. Intake does not investigate devices. A report like that would still need the organization's incident response team to check the machine.
What external action would not change
Once a visitor runs the command, the infection lives on the device. Removing the ad, the hosted page or the payload domains would not uninstall NetSupport, restore stolen browser data or wallet files, or revoke sessions an information stealer may have captured. The mitigations Huntress recommends, such as user awareness, restricting the Run dialog and blocking finger traffic, are endpoint and network controls. External disruption and internal response cover different stages of the same incident.
Trace the Amazon Business lure stage by stage
Choose a view, then select a stage. The attack path follows the sequence Huntress reported. The response view groups the same assets by the party able to act on them. Possible Outtake interventions are hypothetical. This educational section collects no data and links to no attacker infrastructure.
Attack path reported by Huntress
1. Sponsored result
A paid result sits above the real service
Reported evidence
Huntress reports that the top sponsored Google result for "amazon business" at the time of the attacks led to the campaign. Google has since removed the ad.
Possible external intervention
Monitoring paid placements for a protected brand name could support flagging a sponsored result whose destination is not an authorized brand domain or partner.
Control boundary
Huntress does not report how long the ad ran, where it was shown or who saw it. Ad targeting and cloaking can hide a placement from any single monitor.
2. Lookalike host
An Amazon-like label on an unrelated domain
Reported evidence
The ad pointed to a subdomain of an attacker-controlled domain, crafted so the start of the hostname resembled an Amazon orders address, according to Huntress.
Possible external intervention
Comparing the destination host with official brand domains could support evidence that the placement claimed a false affiliation.
Control boundary
A brand-like subdomain on someone else's domain is not a registration the brand can recover. Action depends on the domain's host or registrar.
3. Fake sign-in
A Google Sites page imitates Amazon login
Reported evidence
Visitors were redirected to a counterfeit Amazon sign-in page hosted on Google Sites that asked for an email address or phone number (Huntress).
Possible external intervention
Visual comparison of the page with the brand's real sign-in flow could support an impersonation report to the hosting platform.
Control boundary
Huntress does not say whether the submitted email addresses or phone numbers were stored or reused.
4. Fake CAPTCHA
A verification button copies a hidden command
Reported evidence
Clicking "I'm not a robot" copied a disguised command to the clipboard and told Windows users to paste it into the Run box. macOS visitors received Terminal instructions (Huntress).
Possible external intervention
Capturing the page behavior as evidence could strengthen a report that the page distributes malware, not only that it imitates a brand.
Control boundary
Whether a visitor runs the command happens on their own device, outside any external monitoring.
5. Payload chain
finger starts a layered infection
Reported evidence
The command used the Windows finger utility to fetch a hidden script, which led through IronPython, a loader, Doina and a Node.js package to an alleged Lumma Stealer sample and a pirated NetSupport Client (Huntress).
Possible external intervention
None on the device. External work could support sharing the reported indicators with the security tools that do act on endpoints.
Control boundary
Removing external assets does not uninstall malware, recover stolen data or revoke sessions taken from an infected machine.
6. Infrastructure
Themed domains and a borrowed IP stay online
Reported evidence
Huntress linked barbecue-themed command and payload domains through common hosting, DNS and payloads, and found NetSupport traffic going to an IP that otherwise hosts a Detroit restaurant's website. The infrastructure was still functional at publication (Huntress).
Possible external intervention
Threat graphing could support mapping the related domains so they are reported together instead of one at a time.
Control boundary
Shared hosting alone does not prove common control, and the restaurant IP shows that legitimate assets can sit inside an attacker's path.
Who can act on each asset
Ad platform
The search ad and its advertiser account
Reported evidence
Google removed the ad. Huntress notes that ad-click markers such as aclk and gclid identify sponsored clicks, and that reporting them to Google can cut attackers off from the ad system until they open a new account (Huntress).
Possible external intervention
A report could cite Google's misrepresentation policy with the query, ad text, destination and click identifiers as evidence.
Control boundary
The ad platform decides on removal and account action. A replacement account can return with new ads.
Site hosting provider
The counterfeit sign-in page
Reported evidence
The fake Amazon sign-in page and fake CAPTCHA were hosted in a private Google Sites location (Huntress).
Possible external intervention
An abuse report could package screenshots of the impersonation and the command-copying behavior.
Control boundary
Huntress does not report whether this page was removed. The hosting provider controls that decision.
Domain registrar and host
The redirect domain and payload domains
Reported evidence
The redirect subdomain sat on an attacker-controlled domain, and payloads came from a set of themed and non-themed domains listed in Huntress's indicators.
Possible external intervention
Reports to registrars and hosts could cover the connected set, with the evidence tying each domain to the chain.
Control boundary
Registrars and hosts apply their own abuse policies and evidence thresholds.
Enterprise security team
Infected computers and stolen credentials
Reported evidence
Huntress's recommendations include user awareness, restricting the Run dialog, blocking outbound TCP port 79 and watching for finger.exe and NetSupport in unusual folders (Huntress; Dark Reading).
Possible external intervention
Outtake describes sending brand threat intelligence into SIEM, SOAR and ticketing tools (Brand Protection). That could deliver external indicators to the team running endpoint controls.
Control boundary
Isolation, malware removal, password resets and session revocation remain internal incident response tasks.
Brand owner
The impersonated identity
Reported evidence
The campaign used the Amazon Business name and an imitation Amazon sign-in page. Huntress does not report a compromise of Amazon's own systems (Huntress).
Possible external intervention
A brand team could hold the evidence, coordinate reports across providers and publish guidance through official channels.
Control boundary
The brand owner cannot remove third-party ads or pages directly. It depends on the providers above.
Why sponsored brand searches keep working
A work task becomes the lure
The search in this case was not unusual. Amazon Business is a purchasing service, and looking it up is part of ordinary procurement work. A sponsored result appears above the organic listing for that same query, so a searcher who clicks the first link may never reach the real service. Huntress notes that Google's paid ad function can unwittingly host malicious links, and that each ad click passes through a billing redirect before reaching the advertiser's page. For readers who want the evergreen background, Outtake's explainer on paid search abuse covers how teams separate impersonation from competition.
Trusted hosting lowers suspicion
The counterfeit sign-in page did not sit on a misspelled Amazon domain. It was a page on Google Sites. Huntress's MacSync case showed the same idea with a shared conversation on the real claude.ai domain, posted under the display name "Apple Support." When the hosting platform is large and familiar, checks that only look for newly registered lookalike domains can miss the page entirely. Google's cloaking policy clarification names click trackers that redirect users to malicious sites as prohibited, which gives a report a clear policy basis.
ClickFix moves the decisive step to the user
Microsoft's August 2025 analysis of ClickFix describes the technique as tricking users into running commands themselves, often combined with malvertising and brand impersonation, and frequently delivering Lumma Stealer. Because a person launches the command, Microsoft notes, a campaign could get past conventional and automated security solutions. Huntress reports that ClickFix accounted for about 2.2% of its managed detection signals this year, and nearly 99% of those detections were high severity. In Dark Reading's interview, Reddish said, "The difference here is the depth of the chain."
Signals brand and security teams can watch
Brand teams and security teams see different parts of this campaign. Both views are useful, and neither is complete alone.
For brand, fraud and trust teams, the public-facing signals include:
- Sponsored results for the brand's name or product names whose destination is not an authorized domain or partner.
- Hostnames that place a brand-like label as a subdomain of an unrelated domain.
- Sign-in pages imitating the brand on website builders or other shared hosting platforms.
- Customer or employee reports describing a CAPTCHA or verification step that asks them to paste text into a system prompt.
For security operations teams, Huntress and its researchers point to signals on managed devices and networks:
- Any finger.exe execution on Windows and outbound traffic on TCP port 79, which Huntress says most networks do not block today.
- Run dialog launches of the command interpreter with padded or unusual characters.
- IronPython archives in Temp folders, node.exe running from ProgramData, and NetSupport clients in unapproved locations or on TCP port 725.
- Ad-click markers such as aclk and gclid in browser history or proxy logs, which can tie an infection back to a specific ad campaign.
A ClickFix prompt is a strong signal on its own. As Huntress puts it, no legitimate CAPTCHA asks anyone to paste a line of text into an operating system command prompt.
Limits of this analysis
The incident facts come from one primary investigation. Dark Reading adds interview comments but no separate telemetry. The figure of at least three affected customers reflects Huntress's own visibility, not a total count. The sources do not establish how long the ad ran, how many people encountered it, which countries it reached, whether the advertiser account was new or compromised, what data was stolen, or who operated the campaign. The status of the Google Sites page and the payload domains after publication is also unknown.
The Outtake analysis has separate limits. Whether any monitor would see a particular ad depends on configured brands and keywords, the advertiser's targeting, and cloaking that shows scanners different content. Removal depends on ad platforms, hosting providers and registrars applying their own policies. Outtake's published materials do not claim control over search rankings, and external disruption does not repair infected computers. This article makes no claim about detection or response times for this case.
Frequently asked questions
Was Amazon hacked in this campaign?
No compromise of Amazon's systems is reported. Huntress describes impersonation: a paid search ad, a deceptive hostname on an unrelated domain and a counterfeit sign-in page on Google Sites. The Amazon Business name supplied credibility, not access.
Did removing the ad stop the campaign?
Not by itself. Huntress says Google removed the ad, but the infrastructure behind the later stages remained online and functional when it published. Huntress also notes that attackers cut off from the ad system can return with a new account.
What could a brand team see from outside the victim's network?
The sponsored result, its destination host, the Google Sites sign-in page and the fake CAPTCHA were all public-facing. The finger command, loaders, stealer and NetSupport client ran on victims' computers and were visible only through endpoint investigation like Huntress's.
How can investigators tell that a visit came from a search ad?
Huntress explains that Google's ad-click redirect leaves markers such as aclk in the URL path and gclid in the query string, along with campaign parameters. Huntress says investigators can use those markers in browser history, proxy logs or URL telemetry to reconstruct the ad-to-site chain even after later redirects (Huntress).
What could Outtake not have addressed in this incident?
Outtake's published capabilities concern external assets: ads, domains, hosted pages and connected infrastructure (Brand Protection; Platform). They do not describe removing malware from computers, recovering stolen wallet data or revoking sessions. Those steps belong to the affected organization's incident response.
Sources
- Huntress, Amazon-Themed ClickFix Campaign Serves Up Finger-Licking Malware, October 21, 2026.
- Dark Reading, Decades-Old Windows Utility Powers Modern ClickFix Attack, October 21, 2026.
- Huntress, How a Google Search for Claude Led to MacSync, August 17, 2026 (separate, earlier campaign; context only).
- Microsoft Threat Intelligence, Think before you Click(Fix), August 21, 2025 (technique background).
- Google Ads, Misrepresentation policy and cloaking clarification.
- Outtake, Brand Protection, Platform, Recon Agent and Intake.
- Outtake, What is paid search abuse? and Brand impersonation: detection and removal.