What happened
On November 11, 2026, the threat intelligence team at Stormshield, a French cybersecurity company, published research on a campaign it calls Trigram (Stormshield). According to the report, the operators impersonate about 50 companies, most of them in artificial intelligence, blockchain and Web3, using fake job offers, a fake conference invitation and fake parcel-tracking pages. Every path ends the same way: the target is talked into pasting a command into their own terminal, which installs an information stealer. Stormshield says the campaign targets the AI, blockchain and logistics sectors in Europe and the United States.
Stormshield attributes Trigram with high confidence to North Korea, as a continuation of a cluster Proofpoint tracks as UNK_DeadDrop and of the group known as Kimsuky or APT43 (Stormshield). That is Stormshield's assessment, built on infrastructure, tooling and identities it correlates with earlier research. No government statement accompanies it.
One framing point matters before the details. The named companies are the brands being imitated. Stormshield does not report that any of them was breached or that their real careers sites were altered. The people at risk are the developers and other professionals who receive the offers, and Stormshield does not publish a victim count.
A job offer with a referral code
The most common lure, in Stormshield's description, is an email from a supposed recruiter offering a fully remote software development role with a generous salary. It links to a careers site and supplies a referral code to enter when applying (Stormshield). The sites copy the impersonated company's visual identity and pair its name with a hiring word. Stormshield lists naming patterns such as brand plus council, hires, careers or talents, and prefixes such as buildwith, join and inv2, mostly on low-cost endings like .xyz, .ink, .site, .space and .online. Its examples include buildwithritual[.]xyz for Ritual, joinwebflow[.]xyz for Webflow, mantlehires[.]xyz for Mantle and somniabuilders[.]network for Somnia.
The site behaves like a real application flow. It filters visitors first: only Windows, macOS and Linux desktops get through, and a server-set cookie named cn_ip_restricted blocks some visitors entirely (Stormshield). Applicants who pass are asked for their name, email address, the role they want, a LinkedIn profile or portfolio, and a personal statement. That information reaches the operators before any malware is involved.
The verification step that swaps the clipboard
Next comes a fake anti-spam check. The site asks for the referral code from the email and validates it with the operators' server, which can reject a code as already used or locked to another IP address (Stormshield). That gate matters for defenders as well as victims, because it means the dangerous step is shown only to people holding a valid code.
Once the code is accepted, the page shows a harmless-looking command and says it proves the applicant is not a robot. Stormshield found that the first command only sends a random value back to the server. The page then displays a deliberate fake timeout and asks the applicant to try again. This time the copy button, and even a manual copy, places a different command on the clipboard. That second command downloads a script named verify_w on Windows or verify_m on macOS and Linux, and it ends in a newline so it runs as soon as it is pasted (Stormshield). This pattern of tricking a person into running a pasted command is widely known as ClickFix.
The downloaded script installs a portable copy of Node.js if needed and fetches a JavaScript payload called install_guard.js. Stormshield says that component deploys an infostealer that collects browser credentials, cryptocurrency wallets and files, then installs agents that keep access to the machine (Stormshield). SafeDep, which analyzed the same payload, reports that on macOS and Linux it shows a fake system window asking for the user's password (SafeDep).
Three lures on one server
Recruiting is the largest scenario, but not the only one. One site, icagllmbt[.]space, invites targets to a conference on generative AI and blockchain whose name copies a genuine conference series, and lets them nominate a substitute by forwarding a ticket and access code, which turns the target into a distributor (Stormshield). A third set of pages copied the parcel-tracking interfaces of DHL, DPD and Aramex and, according to Stormshield, appeared on the operators' server only briefly, between October 8 and 10, 2026.
Stormshield links the three scenarios through a shared server, the same verify_w and verify_m scripts and the same code-then-command flow (Stormshield). The conference domain also appears in a separate incident. SafeDep reported on November 3 that a forged commit, force-pushed on October 31 to the GitHub repository for antd-img-crop, an open-source image-cropping component with about 98,000 weekly npm downloads, added a build step that fetched the same scripts from icagllmbt[.]space (SafeDep). SafeDep says the package published on npm was not affected, and it also documented a fake Amber Group careers portal using the same clipboard trick.
A back office built like a marketing platform
The most unusual part of the report is the operators' tooling. The consoles were bundled into the same JavaScript that every decoy visitor receives, so Stormshield could reconstruct them (Stormshield). They track each target step by step, from copying the command to the payload running and a saved system password. They test every decoy domain every 30 seconds so operators notice quickly when one is taken down. They score targets as important, moderate or useless, and they record visitor sessions with a commercial analytics service.
A campaign management application, which its operators call rms-platform, was exposed with debug mode switched on. Stormshield says its route list and configuration names describe the whole campaign lifecycle: building target lists by scraping GitHub, sending mail through Resend, Brevo, Mailgun and Smartlead with daily caps and random delays, tracking opens and replies, issuing the referral codes, and a cryptocurrency wallet module whose purpose Stormshield rates with only low to moderate confidence (Stormshield). It counted 55 Resend keys, about 50 of them prefixed with an abbreviation of an impersonated brand, plus 13 Brevo keys and 3 Smartlead keys. Shared Brevo verification records in the DNS of several lure domains tie groups of them to single sending accounts. Stormshield also found documentation apparently copied from an AI assistant and a configured Gemini API key, which it calls consistent with AI-assisted development.
How it connects to earlier research
Stormshield's attribution rests on overlaps with three earlier reports. Proofpoint described UNK_DeadDrop in June 2026: more than 250 emails to almost 100 organizations in April and May, with fake developer-role offers that spoofed real companies and led to trojanized code repositories. Proofpoint judged the cluster likely North Korean and distinct from Contagious Interview in its own data (Proofpoint). Stormshield says a domain and an IP address that Proofpoint tied to UNK_DeadDrop now host or serve Trigram's back office (Stormshield).
Genians published a report in August 2026 attributing to Kimsuky an operation that used AI tools and Git repositories as command infrastructure. Its indicators include the email address that Stormshield found hard-coded as the console super-administrator, and the same IP address (Genians). Stormshield also ties a database user name in the platform's configuration to a commit author in the GitLab relay that SafeDep documented, and on that basis assesses with high confidence that the antd-img-crop compromise and Trigram share operators (Stormshield).
What public records showed on November 12
To see how much of the lure infrastructure was still standing a day after publication, Outtake Labs checked the 102 lure domains listed in Stormshield's indicator table against public registration (RDAP) records and public DNS on November 12, 2026, at about 13:05 UTC. No lure site was visited.
Registration records returned for 97 of the 102. All 97 were registered between October 11 and October 29, 2026, which fits Stormshield's description of infrastructure built to roll out new brands quickly and absorb the loss of individual domains. Thirty-eight carried a client hold or server hold status, a suspension set by a registrar or registry that takes a domain out of DNS. Forty-three of the 102 did not resolve, including all 38 on hold. The other 59 still resolved. Sixteen of those pointed into the same 193.160.32.x address range as the decoy server Stormshield documented, and 32 pointed to a single address whose role we could not determine. A resolving name is not proof of a live lure page, and records can change after a check. Suspensions were under way for a large minority of the lures, while most of the named domains still resolved.
What is still unknown
Stormshield does not say how many people applied, ran the command or lost credentials or funds. It does not say whether the impersonated companies were notified before publication. The wallet module's purpose is unconfirmed, and the infostealer analysis is still to come. The attribution to North Korea is a vendor assessment based on technical overlaps.
Where Outtake could have helped
This section is a hypothetical analysis, separate from the reported facts above. Nothing in Stormshield's report suggests that any impersonated company uses or lacks Outtake, and Outtake had no part in the research. The question is narrower: at which observable points in this campaign could an external monitoring and response team acting for one of the imitated brands have acted, and where would that help stop?
Outtake's brand protection product lists phishing and lookalike domains, and social and impersonation attacks, as core use cases (Outtake brand protection). Its platform describes a Brand entity that covers fake domains, phishing attempts, impersonation, fake employees and fake job posts (Outtake platform). The fit here is strong for detecting the fake careers sites and mapping the infrastructure behind them. It is weaker for the parts of the chain that happen inside a target's inbox and terminal.
Watch for the brand inside hiring-themed domain names
Trigram's domain names are predictable once the pattern is known: a brand name joined to council, hires, careers or talents, or behind buildwith, join or inv2. A monitoring program that defines the brand, its product names and its short forms, then watches new registrations for those terms beside hiring words, could help surface lures within days of registration rather than after a candidate complains. Background on why new registrations deserve attention is in our guide to newly registered domains.
Timing matters because every lure in Stormshield's table was registered in October. Finding one in its first days, rather than after a candidate complains, shortens the period in which it can collect applications.
Compare the clone with the real careers page
Stormshield says the sites copy each brand's visual identity identically. Outtake describes agentic search that reads logo, layout, image, video and audio rather than matching keywords alone (Outtake brand protection). Applied here, that kind of visual comparison could help confirm that a cheap .xyz domain is presenting itself as a specific company's hiring portal, which is the evidence a registrar or host needs to see. The visitor filtering and code gate are a real limit. A scanner may see the job listing and application form but not the clipboard step, which appears only after a valid referral code.
Treat one lure as a lead to the rest
Outtake says that from a single flagged domain its agents trace shared hosting, registration patterns, analytics IDs, linked social accounts and operator activity, and that pre-staged infrastructure not yet activated can be exposed (Outtake brand protection). Trigram leaves exactly those kinds of traces: clustered hosting, a single registrar for most names, short registration windows, shared Brevo verification codes in DNS and, per Stormshield, one session-recording account across the decoys. A brand that found one fake careers site could use those signals to look for others carrying its name, and the same picture could show other brands that the operators were impersonating from the same platform. Any grouping still needs more than one signal, because shared hosting alone is common.
Report through the routes that clearly apply
Unlike parked lookalikes, these sites deliver malware and collect personal data, so they fall within the phishing and malware categories that ICANN's registrar and registry abuse obligations cover (ICANN). Outtake's platform lists evidence packaging and platform escalation, (Outtake platform), and its brand protection page describes remediation through registrars, hosts, social platforms, app stores and ad networks (Outtake brand protection). A report that shows the brand being copied, the shared infrastructure and the published research could help a registrar, host or email-sending service act on several related names at once. Code platforms such as GitHub and GitLab, and sending services such as Resend and Brevo, decide on their own terms. Stormshield's finding that the consoles test each decoy every 30 seconds suggests the operators expect removals and can move quickly to replacements, so removal of individual domains reduces exposure without ending the operation. Our guide to dismantling fake domain operations covers why the connected infrastructure matters more than one URL.
Follow a Trigram lure from inbox to infostealer
Choose a view, then select a stage. Each stage shows what Stormshield, SafeDep or public records reported, a clearly labelled hypothetical intervention by an external team acting for an imitated brand, and the limit of that intervention. Nothing here is a live scanner or a measured outcome. Malicious domains are written in a safe, non-clickable form.
How the recruiter lure works
Offer
A recruiter email offers a remote developer job
Reported evidence
The email advertises a fully remote software role with a substantial salary, links to a careers site and supplies a referral code (Stormshield). Stormshield says the platform behind it builds target lists by scraping GitHub.
Possible external intervention
Hypothetical: an imitated brand that publishes how its real recruiters make contact, and monitors for its name in hiring-themed domains, could give candidates something to check the offer against.
Control boundary
The email goes straight to the target's inbox. An outside team does not see it unless a recipient reports it.
Clone site
The careers site copies the brand
Reported evidence
Sites such as buildwithritual[.]xyz and joinwebflow[.]xyz copy the real company's visual identity and present a job listing (Stormshield). About 50 brands are imitated.
Possible external intervention
Hypothetical: watching new registrations for the brand beside words like hires, careers and council, then comparing page layout and logo with the real site, could help confirm the impersonation (Outtake brand protection).
Control boundary
Detection depends on the brand's names and short forms being defined. A lure that uses an unrelated name is harder to find by name.
Filter
Only desktop visitors get through
Reported evidence
A script admits Windows, macOS and Linux desktops only, and a server-set cookie named cn_ip_restricted blocks some visitors entirely (Stormshield).
Possible external intervention
Hypothetical: reviewing the site as a desktop visitor would, from more than one network location, could help capture what a real target sees.
Control boundary
Filtering is designed to hide the lure from scanners and researchers. Some visits will see a disabled page.
Application
The applicant hands over personal details
Reported evidence
The form collects full name, email, the role sought, a LinkedIn profile or portfolio and a personal statement, and sends them to the operators' server (Stormshield).
Possible external intervention
Hypothetical: evidence that a site collects applicants' personal data while impersonating a brand strengthens a report to the registrar or host.
Control boundary
Data already submitted cannot be recalled by a takedown.
Code gate
A referral code unlocks the next step
Reported evidence
The site validates the code with the server, which can reject it as already used or locked to another IP address (Stormshield).
Possible external intervention
Hypothetical: documenting the code gate alongside Stormshield's research explains to a provider why a page that looks like a careers form is part of a malware chain.
Control boundary
Without a valid code, an outside review usually cannot reach the malicious step itself.
Clipboard
A fake timeout swaps the command
Reported evidence
A harmless first command is followed by a deliberate timeout. The retry copies a different command that downloads verify_w or verify_m and runs as soon as it is pasted (Stormshield).
Possible external intervention
Hypothetical: none from outside at this moment. This step happens on the target's own machine.
Control boundary
Preventing execution depends on the target recognizing that no hiring process asks for a pasted terminal command, and on their own security tools.
Payload
The stealer installs and reports back
Reported evidence
The script installs Node.js and fetches install_guard.js, which deploys an infostealer for credentials, wallets and files, plus agents that keep access (Stormshield). SafeDep reports a fake system password prompt on macOS and Linux (SafeDep).
Possible external intervention
Hypothetical: shared infrastructure found while investigating the lure, such as related command servers, can be passed to the brand's security team and to providers.
Control boundary
Removing a lure does not clean an infected machine. Incident response belongs to the affected person or their employer.
Where the operation leaves traces
Names
Predictable, hiring-themed domain names
Reported evidence
Stormshield lists patterns such as brand plus council, hires, careers or talents, and prefixes buildwith, join and inv2, mostly on .xyz, .ink, .site, .space and .online (Stormshield).
Possible external intervention
Hypothetical: those patterns can become watch terms for each defined brand.
Control boundary
Operators can change patterns once they see them caught.
Registration
Short registration windows
Reported evidence
On November 12, 2026, registration records returned for 97 of the 102 lure domains in Stormshield's indicator table. All 97 were registered between October 11 and October 29, 2026.
Possible external intervention
Hypothetical: batches of new names registered days apart, through one registrar, are a clustering signal worth checking together.
Control boundary
Timing and registrar alone do not prove a common operator.
Hosting
Clustered hosting and shared verification codes
Reported evidence
Stormshield found the three lure types on one server and shared Brevo verification codes in the DNS of several lure domains (Stormshield). On November 12, 16 still-resolving lures pointed into the same 193.160.32.x range as that server.
Possible external intervention
Hypothetical: tracing shared hosting, registration patterns and analytics IDs from one flagged domain could help find related lures, including other brands' (Outtake brand protection).
Control boundary
Shared hosting is common, so groupings need several signals.
Senders
One sending account per imitated brand
Reported evidence
The platform held 55 Resend keys, about 50 with brand prefixes, plus 13 Brevo and 3 Smartlead keys, with daily caps and random delays (Stormshield).
Possible external intervention
Hypothetical: evidence that a sending account is impersonating a brand can support a report to that email service.
Control boundary
Sending services apply their own policies, and the brand cannot see the operators' accounts directly.
Relay
A GitLab file that points to the command server
Reported evidence
The payload reads its command server address from a public GitLab file. Stormshield saw the relay move to a repository created on November 3, 2026 (Stormshield).
Possible external intervention
Hypothetical: relays and code repositories tied to the campaign can be reported to the code platform with the research that links them.
Control boundary
The operators can repoint a relay by editing one file, so each removal buys time rather than ending the operation.
Status
Suspensions under way, much still reachable
Reported evidence
On November 12, 38 of 97 lure domains with records carried a client hold or server hold status, 43 of 102 did not resolve and 59 still resolved. No site was visited.
Possible external intervention
Hypothetical: re-checking known lures and their neighbors over time could help confirm which reports took effect and which need escalation.
Control boundary
Resolution does not prove a live page, and records change after any single check.
What brand, recruiting and security teams can do now
Several steps do not depend on any vendor:
- Tell candidates how you hire. State on the real careers page which domains and channels recruiters use, and that no step asks candidates to paste a command into a terminal. Add an address for reporting suspected recruiting fraud.
- Search for your name beside hiring words. Look up recent registrations that pair the company or product name with hires, careers, talents, council, join or buildwith.
- Brief your developers. Treat a pasted command or an unexpected build step as a stop sign. The antd-img-crop commit shows the same payload can arrive through a code dependency, not only an email.
Limits of this analysis
The core facts come from one research report, with partial corroboration. SafeDep independently documented the payload, the antd-img-crop commit and the conference domain, and Genians' earlier indicators include the super-administrator address Stormshield found. No independent news coverage of Trigram had appeared when this was written. Our checks cover registration and DNS status for the 102 lure domains, not page content or victim impact.
The limits of a hypothetical Outtake intervention are separate. Outtake cannot stop registrations, does not see lure emails sent to individuals, may not reach steps hidden behind visitor filters and referral codes, and depends on registrars, hosts, code platforms and email services to act on reports. None of this guarantees a timeline or outcome in a specific case. A step-by-step guide to the reporting process is in how to take down a phishing or lookalike domain.
Frequently asked questions
Were the companies named in the Trigram report hacked?
Not according to the report. Stormshield describes them as impersonated brands whose names and visual identity were copied onto fake careers sites. It does not report that any of their systems or real careers pages were compromised (Stormshield).
How can a candidate tell a Trigram-style offer from a real one?
The clearest sign is the request to copy a command into a terminal or PowerShell to prove you are human. Stormshield's analysis shows that step is what installs the malware (Stormshield). Other signals include a careers site on a cheap domain ending that pairs the company name with words like hires or council, and a referral code you must enter to continue. Checking the company's own careers page directly is safer than following an emailed link.
Is Trigram the same as Contagious Interview?
Researchers treat them as related but separate. Proofpoint describes UNK_DeadDrop as likely North Korean and distinct from Contagious Interview in its data (Proofpoint). Stormshield calls Trigram a continuation of UNK_DeadDrop and Kimsuky activity, with new lures such as ClickFix portals and a conference invitation (Stormshield). SafeDep compares the antd-img-crop chain to Contagious Interview on technique only (SafeDep).
Can these fake careers sites be taken down?
Sites that impersonate a brand to deliver malware and collect personal data fall within the phishing and malware categories covered by ICANN's registrar and registry abuse obligations (ICANN). On November 12, 38 of the lure domains we checked already carried a hold status. Stormshield reports that the operators test each decoy every 30 seconds, so individual removals are likely to be followed by replacements (Stormshield).
What could Outtake do in a case like this, and what could it not do?
Hypothetically, monitoring an imitated brand's names in new hiring-themed domains, comparing clones with the real site, tracing shared infrastructure from one lure and packaging evidence for registrars and hosts could help find and disrupt fake careers sites sooner (Outtake brand protection). Outtake cannot see the lure emails in a target's inbox, cannot reach steps hidden behind referral codes without one, cannot clean an infected machine and cannot promise a timeline that depends on providers' decisions.
Sources
- Stormshield, Trigram Campaign, November 11, 2026
- SafeDep, antd-img-crop GitHub Repository Hijacked to Deliver OverLord Stealer, November 3, 2026
- Proofpoint, Don't Fear the Repo: UNK_DeadDrop Phishing Campaign Targets Developers to Steal Cryptocurrency, June 2026
- Genians, Kimsuky Integrates AI into Attack Operations, From AI-Generated Decoy Documents to a Local LLM, August 10, 2026
- RDAP registration records, retrieved November 12, 2026, for the 102 lure domains in Stormshield's indicator table, for example icagllmbt[.]space, buildwithritual[.]xyz and mantlehires[.]xyz
- Public DNS lookups (Cloudflare DNS over HTTPS) for the same 102 domains, checked November 12, 2026 (no lure site was visited)
- ICANN, Advisory: Compliance With DNS Abuse Obligations in the Registrar Accreditation Agreement and the Registry Agreement, February 5, 2024
- Outtake, Brand Protection and Platform, retrieved November 12, 2026
- Outtake articles: What is a newly registered domain (NRD)?, How to take down a phishing or lookalike domain, How to take down fake domains and phishing sites: dismantle the operation, not just the URL