Malvertising attacks now hide on trusted domains: what the fake Claude installer campaign teaches us

Malvertising attacks now hide on trusted domains: what the fake Claude installer campaign teaches us

Last updated: November 19, 2026

Short answer: A malvertising attack uses paid online ads to send people to malware, and the fake Claude installer campaigns show it no longer needs a fake website. Since at least July 2026, researchers have reported sponsored search ads that led to real claude.ai pages, user-generated Artifacts and shared chats dressed up as Anthropic's download page, which handed out SectopRAT on Windows and Atomic macOS Stealer (AMOS) on Mac.

The address bar was genuine; the content wasn't. Below: how the campaigns worked, why domain-level controls miss this, and which control breaks which step. We make dope.security, an on-device secure web gateway, so weigh the product section accordingly. No single layer covers this chain, ours included.

Key takeaways

  • A trusted domain is not trusted content. Both campaigns used ordinary claude.ai sharing features, not a breach of Anthropic's systems (Gen Digital).
  • Two variants: a fake installer delivering SectopRAT (Huntress, Bing ads) and a paste-a-command lure delivering AMOS on Mac (Gen Digital, Google Ads).
  • Domain allowlists and DNS verdicts see claude.ai and pass it. The malicious part is the URL path and what follows the click.

What is a malvertising attack?

A malvertising attack is when criminals buy or hijack online ads, often sponsored search results, to push people toward malware or malicious pages. The ad is the delivery vehicle: the attacker doesn't need to compromise the site where the ad appears, only to buy a spot that looks like the right answer.

Malwarebytes defines malvertising as "the use of online advertising to distribute malware with little to no user interaction required" (Malwarebytes). MITRE ATT&CK (T1583.008) notes such ads can be "prominently placed within search engine results," making it harder "to distinguish between actual search results and advertisements" (MITRE ATT&CK).

The most common malicious ads are sponsored search results for popular software that lead to fake download pages.

How is malvertising different from phishing and typosquatting?

Phishing is sent to you. Malvertising waits for you to search, so it works well on software downloads: the person already expects to run an installer. Typosquatting relies on a lookalike domain. We cover that pattern in why typosquatting is a domain game you can't win. The fake Claude campaigns skipped the lookalike for the first hop: the landing page sat on the real domain.

How did the fake Claude installer campaigns work, step by step?

The reports describe related but different campaigns, so each detail is labeled with its source.

Step 1: The ad

  • Huntress (Windows, Bing): a Bing search for "CLAUDE DESKTOP APP" returned fake download pages plus a sponsored link to the real claude.ai domain (Huntress, July 22, 2026). BleepingComputer reported the campaign on July 23, 2026 (BleepingComputer).
  • Gen Digital (Mac and Windows, Google): the links carried Google Ads tracking parameters, and Gen found several Artifacts tied to two ad campaigns, "a funded and rotating operation" (Gen Digital, October 26, 2026).

Step 2: The landing page on claude.ai

Attackers used Claude's public sharing links to copy the look of Anthropic's download page.

  • Huntress: a public Artifact at claude.ai/public/artifacts/…, with 7,100 page views before Anthropic removed it.
  • Gen Digital: public Artifacts at claude.ai/artifact/… and Shared Chats at claude.ai/share/….
  • Zip Security: claude.ai/public/artifacts/… links promoted through paid search (Zip Security, July 27, 2026).

The main visible warning, per Gen, was a thin banner reading "Content is user-generated and unverified." Neither campaign required "compromising Anthropic's infrastructure," Gen notes.

Step 3a: The download variant (Huntress)

Clicking "Download" on the Huntress Artifact redirected the user off claude.ai, first to claude.ai.download-app[.]us (registered May 9, 2026) and then to downloading-api.it[.]com, which served ClaudeDesktop.exe. Huntress found the executable was actually a JetBrains Chromium component that loaded a malicious libcef.dll placed next to it. A copy named DockerDesktop.exe was written to disk as a scheduled task for persistence, and the loaders used VMProtect packing and graphics-hardware checks to avoid analysis.

Step 3b: The ClickFix variant (Gen Digital)

Gen's Mac page offered no installer at all. A "Quick install via Terminal" box walked visitors through four steps: open Terminal, click Copy, paste, press Enter. The copied text was an obfuscated one-liner. A separate "Install Claude on Windows" Artifact gave three keyboard steps: Win + R, Ctrl + V, Enter. The resulting script pulled its next stage from a github.io page, so a trusted name carried the payload again.

This pattern is called ClickFix. As Gen puts it, "Nothing suspicious is really 'downloaded.'" You copy some text and run it yourself.

Step 4: The payload

  • SectopRAT (Huntress, Windows). A remote access trojan and infostealer, also known as ArechClient2, active since 2019 per BleepingComputer. It targets passwords, card data, files, browser logins and cookies, and messaging apps, and includes hidden remote desktop (HVNC). It looked up its command-and-control address through BNB Smart Chain transactions, a technique called EtherHiding.
  • Atomic macOS Stealer, or AMOS (Gen Digital, Mac). It shows a fake macOS password prompt, then harvests browser passwords and session cookies, Keychain keys, crypto wallets, and files. It can also replace Ledger and Trezor apps with modified copies that ask for the wallet recovery phrase.
  • Gen's Windows payload is a heavily obfuscated .NET executable. Gen doesn't name it SectopRAT, so neither do we.

Step 5: The scale

Huntress reports at least 29 organizations were hit on July 21 and 22, 2026, and that incidents tied to the same redirect domain continued through August (Huntress). Anthropic removed that Artifact; no source we found says the underlying feature abuse is fixed.

Why do DNS filtering and domain allowlists miss trusted-domain abuse?

DNS and domain-reputation controls judge the hostname, and claude.ai is one many companies explicitly allow. Zip puts it plainly: "A trusted domain does not mean trusted content."

The malicious part lived in the path (/artifact/…, /share/…, /public/artifacts/…) and in what came next. The path sits inside the encrypted HTTPS request, so a control has to see the full URL, not just the domain lookup. We go deeper on that difference in DNS filtering vs. URL filtering and SSL inspection and what Cisco Umbrella DNS logs can and can't show about URL paths.

Two caveats. In the Huntress chain, the file came from separate redirect domains, which can be judged on their own. And in Gen's ClickFix chain, nothing was downloaded through the browser at the key step; the user ran a command. The human action is the gap.

Which controls break which step of a malvertising attack?

Step Control What it does, and its limits
Ad click User habit: never install software from a sponsored result; type the vendor address or use a bookmark Gen, BleepingComputer, and Zip all give this advice. Gen names claude.com/download as the real Claude download page.
Landing page on a trusted domain URL-path policy Lets you treat user-generated paths on a trusted domain like a file-share link. It applies to legitimate links on those paths too.
Redirect to the download domain URL categories, including Newly Registered and Dynamic DNS Helps when the domain is categorized as malicious or is new. Huntress's redirect domain was registered May 9, more than 30 days before the July attacks, so a "newly registered" check alone wouldn't have matched it.
Downloaded installer Malware checks on downloads Can catch files with a known-bad verdict. A brand-new or repackaged file may not have one yet.
Pasted command (ClickFix) User habit; endpoint detection and response (EDR) Gen: "Never paste into Terminal (or PowerShell) a command that you did not write yourself." A web gateway can only act on the follow-on web requests, and only if those destinations are categorized.
After execution EDR, isolation, credential resets, session revocation Gen advises resetting important passwords from a clean device, logging out of all sessions, turning on MFA, and moving crypto to a wallet with a new recovery phrase.

Each row reduces exposure; none stops the campaign alone.

How does on-device inspection help with trusted-domain malvertising?

Path-level rules only work if something can see the path, and the path is inside HTTPS. With dope.security, the proxy runs on the endpoint: SSL inspection, URL filtering, and Cloud App Control run on the device, not in a stopover data center (dope.swg). (See cloud proxy vs. on-device SWG.)

  • Landing page (step 2). Admins can block or warn on specific URL paths, such as claude.ai/share/ or /artifact/ links (release notes; custom categories). This is an option, not a default recommendation. Warning shows a page users can click through; Block stops access. A path rule applies to everyone in that policy, including legitimate Artifact and shared-chat links; exceptions can be set per user or group. It doesn't tell a malicious Artifact from a harmless one.
  • Redirect and payload domains (steps 3a and 4). Known-malicious and suspicious destinations are categorized and can be blocked (automatic updates). Newly Registered (domains registered in the last 30 days) and Dynamic DNS categories have been available since November 21, 2023 (release notes), and you can warn on, block, or log them.
  • Download (step 3a). The endpoint checks downloads against malware verdicts. It hashes each download, checks a local cache, and sends unknown hashes to dope.cloud for a verdict (whitepaper). Known-malicious downloads are blocked and logged with the file name, hash and the process that fetched them (SIEM integration).
  • ClickFix (step 3b). This is outside what a web gateway can see. At most, the categories above can block the lure or payload domains a pasted command reaches out to, if those domains are categorized. That's not a guarantee.

None of this replaces EDR, which watches what code does after it runs. For where a web gateway's job ends on downloads, see can Cisco Umbrella stop a malware download?

Checklist: how do you reduce exposure to trusted-domain malvertising?

  • Publish approved download links for AI apps (claude.com/download for Claude), and tell staff to skip sponsored results.
  • Teach one rule: never paste a command from a web page into Terminal, PowerShell, or the Run box.
  • Review web policy for user-generated paths on trusted domains (artifact, share, and public file-host links), not just whole domains.
  • Make sure HTTPS inspection is on where you rely on path rules.
  • Turn on malware checks for downloads, and warn on or block newly registered and dynamic-DNS destinations.
  • Confirm EDR is enforcing on every managed device.
  • Write down response steps: isolate, reset passwords from a clean machine, revoke sessions.

Frequently asked questions

What is a malvertising attack?

A malvertising attack uses online ads, often sponsored search results, to push people toward malware or malicious pages. The attacker doesn't need to breach the site showing the ad. In the fake Claude campaigns, the ads led to real claude.ai pages that users had created to look like Anthropic's download page.

How does malvertising work?

An attacker buys an ad matching a common search, such as a software download. The ad leads to a legitimate-looking page that offers a fake installer or asks the visitor to run a command. Once it runs, malware such as an infostealer collects passwords, cookies, and files.

How do you stop malvertising?

No single control stops it. Combine user habits (never install from sponsored results, never paste commands from web pages), URL-path and category policy at the web layer, malware checks on downloads, and endpoint detection and response on every device. Then prepare response steps for when something gets through.

Can a legitimate website like claude.ai be used to spread malware?

Yes. Gen Digital found fake "Download Claude" pages hosted as public Artifacts and Shared Chats on claude.ai, created with ordinary sharing features. Anthropic's systems weren't breached. A real domain tells you where content is hosted, not who wrote it or whether it's safe.

What is SectopRAT?

SectopRAT, also known as ArechClient2, is a remote access trojan and infostealer that BleepingComputer reports has been active since 2019. Huntress found it at the end of the Bing-promoted fake Claude Desktop campaign in July 2026, where it targeted passwords, cookies, card data, and messaging apps.

What is Atomic Stealer (AMOS)?

Atomic macOS Stealer, or AMOS, is Mac malware that steals passwords, browser sessions, Keychain data, crypto wallets, and files. Gen Digital found it delivered through fake claude.ai install pages promoted with Google Ads. It can also replace Ledger and Trezor apps with copies that steal the recovery phrase.

Does DNS filtering stop trusted-domain malvertising?

Not on its own. DNS filtering judges the domain, and claude.ai is legitimate. The malicious part was the URL path and the follow-on download or command. DNS can still block a redirect or payload domain if it's already known to be malicious.

The domain was real. The content wasn't.

The fake Claude campaigns show where malvertising is heading: paid ads, a trusted domain, and a page anyone could publish. Domain trust alone can't sort that out. Controls need to look at the path, the download, and what runs afterward.

If you want to see how path-level policy and download checks work on the endpoint, email sales@dope.security.

Secure Web Gateway
Secure Web Gateway
← back to blog Home