What happened
On October 26, 2026, Gen Digital, the consumer security company behind Norton and Avast, published research describing two malware campaigns that used Anthropic's own claude.ai domain as the landing page for paid Google ads (Gen Digital). People who clicked the ads reached pages that carried Claude branding and explained how to "install" Claude. The pages offered no installer. They walked visitors through copying a command into Terminal on a Mac, or into the Run dialog on Windows, and that command started a malware infection.
The pages were not hosted on a lookalike domain. According to Gen, the attackers built them with two ordinary Claude sharing features, Shared Chats and public Artifacts, which let any user publish content at an address on claude.ai. Gen states that neither campaign required compromising Anthropic's infrastructure or registering an imitation domain (Gen Digital).
Two campaigns with the same front door
Gen describes a macOS campaign and a separate Windows campaign. In both, the links it analyzed carried Google Ads tracking parameters, which Gen treats as evidence that the operators paid to place the pages at the top of search results. Gen says it found several different Artifacts in the same cluster, connected to two different ad campaigns, and calls the activity "a funded and rotating operation, not a single case" (Gen Digital).
The macOS lure appeared both as a public Artifact styled as a "Download Claude" page and as a Shared Chat presented as a guide to running Claude Code on a Mac. Gen's side-by-side comparison shows the Artifact reusing the Claude logo, a near-identical headline and a layout close to Anthropic's genuine download page at claude.com/download. The Windows lure was an Artifact titled "Install Claude on Windows" (Gen Digital).
Both pages displayed Anthropic's own warnings. Artifact pages carried a thin banner reading "Content is user-generated and unverified," and the Shared Chat carried a disclaimer that its content "may include unverified or unsafe content that does not represent the views of Anthropic." Gen's point is that the domain, certificate and branding still looked genuine, and the lure sat directly beneath those notices (Gen Digital).
What the pasted command did
Gen classifies both lures as ClickFix, a social engineering pattern in which victims run the attacker's command themselves. On macOS, the page showed a "Quick install via Terminal" box with four steps: open Terminal, click Copy, paste and press Enter. The copied text was an obfuscated one-line command that hid the attacker's server address inside an encoded string (Gen Digital).
Running it started a chain that Gen traced through a loader script, a second script and a main payload from the Atomic macOS Stealer (AMOS) family. Gen reports that the stealer waited for Full Disk Access, showed a fake macOS password prompt that would not close until it received a working password, then collected browser passwords and session cookies, Keychain material, cryptocurrency wallets and files with extensions such as .pdf, .docx, .pem, .seed and .ovpn. It packed the data into a ZIP archive, uploaded it to the attacker and deleted its temporary files (Gen Digital).
The report describes a second stage of theft aimed at hardware wallet users. When the stealer found Ledger Live, Ledger Wallet or Trezor Suite, it replaced the application with a trojanized copy. Later, the fake application asked the user to "restore" or "validate" a recovery phrase and sent the words to a separate attacker server built to collect seed phrases (Gen Digital).
On Windows, the page asked visitors to press Win + R, paste with Ctrl + V and press Enter. The pasted PowerShell command displayed reassuring progress messages while it fetched a next stage from a github.io page, which delivered a packed and encrypted .NET payload. Gen says that payload injected itself into a normal Windows process, checked which antivirus was installed, captured screenshots, read active window titles, kept a connection open to an attacker server and searched for desktop and browser-extension cryptocurrency wallets. Gen describes that as the behavior of a remote-access tool (Gen Digital).
What is new and what is not
Paid ads that point at user-generated pages on AI platforms are not new. In December 2025, Kaspersky described Google ads leading to a shared ChatGPT conversation that posed as an installation guide for OpenAI's Atlas browser (Kaspersky), and Huntress traced an AMOS infection to highly ranked ChatGPT and Grok conversations in search results (Huntress). In June 2026, Trend Micro reported a Google Ads campaign that moved from GitLab Pages to claude.ai Shared Chats in May, with at least 61 share IDs and 33 new ad campaign IDs observed in its final wave. Trend Micro said that after it notified Anthropic, the company banned the responsible accounts, disabled the conversations and began adding abuse mitigations (Trend Micro). Push Security says it has tracked a technique it calls LLMshare since May 2026, in which attackers promote shared chats and Artifacts on ChatGPT and Claude through search ads (Push Security via BleepingComputer). Separately, on October 9, Push described a different chain in which a Google ad for "claude mac" showed bing.com as its destination and redirected visitors through a compromised website to a lookalike download domain (Push Security).
Gen's October 26 report adds detail on two campaigns within that pattern: a macOS chain that ends in AMOS plus wallet application replacement designed to capture recovery phrases, and a parallel Windows campaign that staged its payload on github.io and delivered a .NET remote-access payload. This article does not treat Gen's findings as a continuation of the Trend Micro campaign. Gen does not link the operators, and the two reports describe different payloads.
What remains unknown
Gen's post does not state a start date or duration for either campaign, or say how many people clicked them or how many devices were infected. It does not say whether it reported the pages to Anthropic or the ads to Google, or whether either has since been removed. It does not attribute the campaigns to a named operator or say whether one group ran both. Reports by GBHackers and Cybersecurity News on October 27 restate Gen's findings without adding independent telemetry. The incident facts here rest on Gen's single primary disclosure, and the other reports serve as context only.
Where Outtake could have helped
This section is a hypothetical analysis. Nothing in Gen's report suggests that Anthropic, Ledger, Trezor or any affected user uses or lacks Outtake, and Outtake played no part in the research. The question is which stages of this campaign were visible from outside, and which of Outtake's published capabilities map to them.
Outtake's Brand Protection page says detection covers domains, social platforms, mobile app stores and paid advertising networks, across text, images, video and audio, and that remediation runs through integrations with registrars, hosts, social platforms, app stores and ad networks. Its Platform page lists ads and code repositories among the surfaces it watches, describes threat graphing that maps adversarial networks, and names evidence packaging, platform escalation and verified takedown as remediation steps.
Watching the ads, not only the domains
Many brand monitoring programs start with lookalike domains. This campaign had none. The visible signal was a sponsored search result that used a software brand's name and logo to advertise a download the brand never published. Monitoring that reads paid search results for a brand's product names, and compares each advertised destination with the brand's known download paths, could have helped surface an ad that pointed at a shared chat or Artifact instead of the official download page. Visual matching matters because the lure copied the logo and layout of the genuine page (Gen Digital). For background on the category, see what is paid search abuse? and our guide to cloned malicious ads. A related live example is the fake Amazon Business ad that led to a ClickFix chain, which used an off-platform lookalike page instead of a trusted host.
Treating a trusted domain as a destination, not a verdict
Allowlists and reputation checks tend to treat a major platform's domain as safe. Reviewing this campaign calls for a different rule: a page on a trusted domain still has an author, and the relevant question is whether the brand published it. In Gen's comparison, the distinguishing details were the URL path and the user-generated content banner (Gen Digital). A triage workflow that classifies landing pages by who controls the content, rather than by which domain serves it, could support faster decisions on this pattern.
Connecting rotating pages into one operation
Gen found several Artifacts tied to two ad campaigns, and Trend Micro's earlier report shows how quickly one operator rotated dozens of share IDs and ad campaigns (Trend Micro). Removing one page at a time leaves the rest running. Outtake's published approach is to map the network behind a single signal by linking shared hosting, registration patterns, analytics IDs and connected accounts (Brand Protection). Applied here, the useful links would be repeated page templates, shared lure text, common command structures, the github.io staging page and the ad identifiers carried in the URLs. Grouping those could let one report describe the whole cluster.
Routing evidence to the parties that can act
No single party could remove every piece of this chain. Google can suspend advertisers, and its policy treats intentional malware distribution as an egregious violation that leads to suspension upon detection (Google Ads policy). Anthropic controls the Shared Chats and Artifacts. GitHub controls the github.io staging page. A brand protection team that assembles one evidence set, covering the ad, the landing page, the copied command's behavior and the staging host, and routes the relevant part to each provider could help shorten the time each piece stays live. Outtake describes these steps as evidence packaging and platform escalation (Platform).
Watching for abuse of wallet brands
The Ledger and Trezor element is harder to reach from outside, because malware installed the trojanized applications locally instead of offering them in a public store. External monitoring could still look for public pages, ads or social posts that offer Ledger or Trezor downloads, or recovery phrase "validation," outside official channels. Gen's report does not say whether any public distribution of that kind existed in this case, so this is a general control rather than a claim about this campaign.
Follow the fake Claude install path stage by stage
Choose a view, then select a stage. The first view follows the attack path described in Gen Digital's October 26 report. The second groups the same campaigns by the party able to act. Reported evidence comes from the cited sources. Possible interventions are hypothetical and do not describe what happened in this case. This section links to none of the malicious pages and collects no data.
Reported attack path
1. Sponsored ad
A paid search ad sends Claude searchers to claude.ai
Reported evidence
Gen found Google Ads tracking parameters (gclid and gad_campaignid) in the links it analyzed and several Artifacts tied to two different ad campaigns, which it calls a funded and rotating operation (Gen Digital).
Possible external intervention
Monitoring paid search results for product names could flag an ad whose destination is a shared chat or Artifact rather than the brand's official download page.
Control boundary
Ads are targeted by region and keyword, so a monitor may not see them. Gen does not report how long the ads ran or how many people saw them.
2. Landing page
The ad lands on a Shared Chat or public Artifact on claude.ai
Reported evidence
The pages were user-generated content on claude.ai with Claude branding and Anthropic's unverified-content notices. Gen says no Anthropic infrastructure was compromised (Gen Digital).
Possible external intervention
Classifying the page by who authored it, then capturing the title, branding reuse, URL path and banner as evidence, could support a grouped report to Anthropic.
Control boundary
Only Anthropic can remove the pages, under its own policies. Gen does not say whether it reported them.
3. Paste
The page tells the visitor to paste a command
Reported evidence
On macOS, four steps led to pasting an obfuscated command into Terminal. On Windows, the steps were Win + R, Ctrl + V and Enter, which ran PowerShell (Gen Digital).
Possible external intervention
Recording the visible instructions and the shared command structure could help link separate pages to the same operation in one investigation.
Control boundary
External monitoring cannot stop a user from pasting a command. That control sits with endpoint security and user training.
4. macOS stealer
AMOS takes passwords, cookies, wallets and files
Reported evidence
The command fetched a loader, a second script and an AMOS payload that waited for Full Disk Access, showed a fake password prompt, then collected browser data, Keychain material, wallets and selected files before uploading a ZIP archive (Gen Digital).
Possible external intervention
Where investigators identify a delivery or upload server, reporting it to its host alongside the landing page evidence could support disruption.
Control boundary
Data that already left a device stays with the attacker. Gen does not publish victim counts, and this article does not assume any.
5. Wallet swap
Trojanized Ledger and Trezor apps ask for recovery phrases
Reported evidence
When AMOS found Ledger Live, Ledger Wallet or Trezor Suite, it replaced the app with a trojanized copy that later asked the user to restore or validate a recovery phrase and sent it to a separate seed-collection server (Gen Digital).
Possible external intervention
Wallet brands could watch public channels for downloads or recovery phrase "validation" offers outside official sites. A seed-collection server, if identified, could be reported to its host.
Control boundary
The replacement happened locally on infected machines, out of external view. Funds taken with a recovery phrase are generally not recoverable.
6. Windows variant
A github.io stage delivers a .NET remote-access payload
Reported evidence
The Windows Artifact, "Install Claude on Windows," ran PowerShell that fetched a next stage from a github.io page and delivered a .NET payload with screenshot, window-title, antivirus-check and persistent server connection behavior, plus wallet searches (Gen Digital).
Possible external intervention
Reporting the github.io staging page to GitHub together with the linked Artifact and ad evidence could help remove a shared dependency of the chain.
Control boundary
Gen does not name the staging page publicly. Removal is GitHub's decision, and devices already infected still need endpoint response.
Who could act at each point
Google Ads
Google Ads, which carried the sponsored results
Reported evidence
The analyzed links carried Google Ads parameters. Google's policy treats intentional malware distribution as an egregious violation that leads to suspension upon detection (Google Ads policy).
Possible external intervention
An ad report that includes the landing page evidence and the campaign identifiers in the URLs could make the violation easier to confirm.
Control boundary
Google reviews reports on its own timeline. Operators can open new campaigns, as Trend Micro observed in an earlier claude.ai campaign (Trend Micro).
Anthropic
Anthropic, which operates Shared Chats and Artifacts
Reported evidence
The lures were hosted with claude.ai sharing features. In June, Trend Micro reported that Anthropic banned accounts and disabled malicious shared conversations after notification (Trend Micro).
Possible external intervention
A grouped report listing every related shared chat and Artifact, with matching templates and commands, could help remove the cluster together.
Control boundary
Gen does not say whether Anthropic was notified about these campaigns or what action followed. Enforcement decisions belong to Anthropic.
GitHub
GitHub, which hosted the Windows staging page
Reported evidence
Gen reports that the Windows PowerShell command reached a github.io page that delivered the encrypted .NET payload (Gen Digital).
Possible external intervention
A report connecting the staging page to the Artifact lure and the ad could support removal of that dependency.
Control boundary
Gen does not publish the page address, so outside parties may not be able to identify it from the report alone.
Brand owners
Brands named in the lures and payloads
Reported evidence
The lures used Claude's name and logo, and the macOS payload impersonated Ledger and Trezor wallet applications (Gen Digital).
Possible external intervention
Brand monitoring across ads, web pages and social posts could surface public lures using these names, and customer warnings could point users to official download paths.
Control boundary
Locally installed fake wallet apps are invisible to external monitoring. Monitoring cannot confirm intent without the page content.
Security teams
Enterprise security teams protecting employee devices
Reported evidence
Gen's guidance is never to paste a command you did not write and to download software only from the vendor's site (Gen Digital).
Possible external intervention
Security teams could warn staff that software should come from official sites or managed catalogs, and share the lure pattern internally.
Control boundary
Endpoint detection and incident response are outside Outtake's published scope. This article makes no claim about any organization's exposure.
What this approach would not have solved
External monitoring sees what is public. It would not have shown which users clicked, what ran on their machines or what data left them. Gen documented the infection chain from samples. Reaching the same understanding inside a company requires endpoint telemetry and incident response, which sit outside Outtake's published scope.
Removal depends on platforms. Google, Anthropic and GitHub make their own enforcement decisions under their own policies and timelines. A clear evidence package can make a report easier to act on, but it cannot compel action. Cloaking also limits outside visibility. Push Security's Adception research shows malicious pages that appear only when visited through a specific ad path (Push Security), and similar filtering would hide pages from scanners that do not reproduce that path.
Ad visibility varies by market and query. Search ads are targeted by location, language and keyword, and monitoring that does not query from the targeted regions with the targeted terms can miss an ad entirely. Trend Micro's earlier report found traffic concentrated in specific regions (Trend Micro).
Some of the harm cannot be reversed. Once someone types a recovery phrase into a trojanized wallet application, removing a web page does not return the funds. User habits still matter, and Gen's own advice is never to paste a command you did not write and to download software only from the vendor's official site (Gen Digital).
The incident facts in this article come from one primary disclosure. If Anthropic, Google or other researchers publish more detail, parts of this analysis may need revision.
Frequently asked questions
Was Anthropic's infrastructure hacked?
No, according to Gen. The report says neither campaign required compromising Anthropic's infrastructure or registering an imitation domain. The attackers used Shared Chats and public Artifacts, features available to ordinary Claude users (Gen Digital).
How could someone tell the page was not an official download?
The pages displayed Anthropic's user-generated content notices, and their addresses pointed to a shared chat or Artifact rather than Anthropic's download page, which Gen identifies as claude.com/download. An instruction to paste a command into Terminal, PowerShell or the Run dialog to install an app is also a warning sign (Gen Digital).
Is this the same campaign Trend Micro reported in June?
Gen does not say so. Trend Micro's June report described claude.ai Shared Chat abuse that delivered the MacSync infostealer (Trend Micro). Gen describes AMOS on macOS, a separate .NET remote-access payload on Windows and the use of public Artifacts. Without attribution from either researcher, this article treats them as separate reports of the same technique.
Were the ads and pages taken down?
Gen's post does not say, and this article does not assume either outcome. Google's ad policy states that intentional malware distribution leads to account suspension upon detection (Google Ads policy).
Is Outtake involved?
No. Outtake had no role in Gen's research, and this article describes no response by Outtake. Outtake has separately published work about an investigation agent it built on Claude. The section on where Outtake could have helped is hypothetical and based on Outtake's published Brand Protection and Platform capabilities.
Sources
- How Criminals Abuse Claude's Own Website to Spread Malware, Gen Digital, October 26, 2026 (primary disclosure)
- Threat Actors Abuse claude.ai Shared Chat for ClickFix Malvertising Campaign, Trend Micro, June 17, 2026 (context)
- The AMOS infostealer is piggybacking ChatGPT's chat-sharing feature, Kaspersky, December 9, 2025 (context)
- AMOS Stealer Exploits AI Trust: Malware Delivered Through ChatGPT and Grok, Huntress, December 9, 2025 (context)
- ClickFix has never been bigger, and AI tools are its most convincing lure yet, Push Security on BleepingComputer, October 21, 2026 (context, sponsored article)
- Adception: malvertising another search engine's search results to redirect to a malicious page, Push Security, October 9, 2026 (context, separate technique)
- Abusing the ad network, Google Ads policy (context)
- Hackers Use Google Ads to Push Fake Claude Downloads Hosted on Real Claude.ai Domain, GBHackers, October 27, 2026 (restates Gen)
- Hackers Abuse Claude's Official Website and Google Ads to Spread Malware, Cybersecurity News, October 27, 2026 (restates Gen)
- Brand Protection and Platform, Outtake (capability descriptions)