Ghostdrive

News · Academy · November 19, 2026

A fake Claude download on claude.ai: how to check a page

In July, ads sent people to a fake Claude Desktop download hosted on claude.ai itself. Why the domain was not enough, and five checks.

DRIVE 369 DAO · 4 min read

In July 2026, attackers paid for Bing ads that sent people searching for "Claude Desktop app" to a page on claude.ai itself: a public, user-made Claude Artifact with a download button that led to malware. Huntress, a security firm, says at least 29 organisations installed a remote-access trojan on 21 and 22 July. The address bar showed a real domain; the page was not the vendor's. The story is circulating again this week, and the lesson holds: a trusted domain tells you who runs the site, not who wrote the page.

What happened

StepWhat Huntress reportsTier
The adA Bing search for "CLAUDE DESKTOP APP" showed sponsored results, including one that led to a public Claude Artifact on claude.aiReported by Huntress
The pageThe artifact showed the note "Content is user-generated and unverified" and had 7,100 page views before it was removedReported by Huntress
The buttonDownload redirected to a lookalike domain registered on 9 May 2026, then to a Windows downloadReported by Huntress
The fileClaudeDesktop.exe was a component of a legitimate JetBrains helper, used to side-load a malicious file. A second file, DockerDesktop.exe, re-infected the machine through a scheduled taskReported by Huntress
The payloadSectopRAT, a remote-access trojan; Huntress found hidden remote desktop (HVNC) in itReported by Huntress
The countAt least 29 organisations, seen as unusual installs, Defender exclusions and odd persistence by Huntress's own security teamReported by Huntress
The responseHuntress reported the artifact to Anthropic; by 22 July it had been removedReported by Huntress

Source: Huntress, published 22 July 2026. It is a vendor's own research, and the article is hedged where it is unsure: it calls the payload a fork of a known family and links the operator to an April 2026 campaign as an analyst's inference, not a confirmed attribution. We have not repeated its indicator list; it is in the article.

Why the domain check failed

"Look at the address" is the standard advice, and here it pointed the wrong way. Any service that lets users publish pages, documents or files puts strangers' content on a domain you trust: artifacts, shared docs, code hosting, forums, form builders. Anthropic's site labelled the page as user-generated, and that label was the only visible sign. Huntress's own advice is the same: do not trust a top-level domain on its own.

Do this week

  1. Get installers from the vendor's own site, typed or bookmarked, or from your system's store or your company's software centre. Not from a search ad, even a sponsored result that looks right.
  2. Read the page, not just the address. On a domain that hosts user content, a note that the content is user-generated means a stranger wrote it, and its download button is a stranger's.
  3. Check where the file came from, not only the file. A real vendor publishes a checksum or a signed package. A single file that looks legitimate is not enough: this chain used a legitimate helper to load the bad one.
  4. IT teams: watch what fresh installs do. The signs Huntress saw were new executables named after popular apps, exclusions added in Defender and new scheduled tasks. Alert on those, and prefer an allow-list for software installs.
  5. If someone ran it, disconnect the machine, change passwords and keys used on it from a different, clean device, remove the scheduled tasks and treat it as a full compromise. A remote-access trojan lets the attacker act as the user.

Where Ghostdrive stands

Ghostdrive's Send and Guard warns before a key is pasted into an AI chat; it does not check downloads. Our desktop builds are listed on the Download page with a SHA-256 for each file, and the page says the builds are not yet signed, so compare the checksum before you run one. A checksum shown on a page you reached through an ad proves nothing, so go to the site directly.

Limits

We read Huntress's report and did not analyse the malware ourselves. Huntress counted only organisations its own team saw, and says nothing about people outside its customers. We did not verify the later reports of similar fake Claude installers that appeared this autumn, so we do not rely on them.

Academy: where AI keys should live · Ten guardrails every AI agent needs · Before you connect an MCP server

Questions

Was claude.ai hacked?
Huntress does not say so. The page was a public Claude Artifact, a page any user can publish, and the site showed the note that the content is user-generated and unverified. Anthropic removed it after Huntress reported it.
How many people were affected?
Huntress says at least 29 organisations fell victim, counted from what its own security team observed, so it may not be the total. The artifact had 7,100 page views before removal.
Is a valid signature on the installer enough?
No. Huntress found the fake used a program component that belongs to a legitimate JetBrains helper to load a malicious file. Check where the download came from, not only the file.