Latest / Decrypted: The UK Cyber Briefing / The fake Claude app that lived on claude.ai, and the 29 firms it caught out
Transcript
- 0:03This is Decrypted. Friday, the twenty-fourth of July.
- 0:07The fake Claude app that lived on claude dot ai, and the 29 firms it caught out.
- 0:15Search for "Claude desktop app" on Bing this week and you could have landed on a download page hosted on Ant
- 0:23hropic's own domain, claude dot ai.
- 0:26It was built from a legitimate product feature.
- 0:30And it was stuffed with an information-stealing trojan.
- 0:33Researchers at Huntress found that between 21 and 22 July, at least 29 organisations installed what they tho
- 0:44ught was Anthropic's Claude Desktop client.
- 0:46They got SectopRAT instead.
- 0:49That's a remote access trojan that harvests browser passwords, cookies, autofill data and card details.
- 0:58The page had been viewed over 7,100 times before it came down.
- 1:03Nobody's domain was spoofed.
- 1:06That is exactly the point.
- 1:08Next. A trusted address, an untrusted page.
- 1:13The mechanism is simple once you see it.
- 1:16Claude Artifacts is a real Anthropic feature.
- 1:20Anyone can publish an interactive page at a public artifacts address on claude dot ai and share the link.
- 1:29It's useful for demos, prototypes and small tools.
- 1:32The attackers used it to publish a page that looked like the official Claude Desktop download.
- 1:39They bought a sponsored slot on Bing, above the genuine search results.
- 1:44Then they let the browser's address bar do their social engineering for them.
- 1:49The installer itself was a legitimate, signed JetBrains component, repurposed to sideload a malicious D L L.
- 1:58It used VMProtect packing and G P U-based checks to dodge sandboxes.
- 2:04Every link in that chain was trustworthy alone.
- 2:08Chained together, none of it was.
- 2:10And finally. The Secure by Design lesson.
- 2:15The failure sits with a design choice made long before this campaign.
- 2:20That choice was allowing user-published content to live on the same origin as the official product.
- 2:27If you judge safety by the domain name, and plenty of people who know better do, you can't tell a shared prot
- 2:34otype from an official release when both sit under claude dot ai.
- 2:38The established fix comes from platforms that learned this lesson the hard way.
- 2:43It's origin separation. You serve user content from a distinct domain that can't borrow the parent brand's re
- 2:51putation. Google keeps user uploads off google dot com.
- 2:55GitHub Pages sits on github dot io rather than github dot com.
- 3:00Anthropic pulled the page once Huntress reported it.
- 3:04But the design gap that let it exist is still there.
- 3:07For UK organisations, this lesson goes well beyond one AI vendor.
- 3:13Any product that lets customers or the public post content under your primary domain inherits the same exposu
- 3:20re. Think of a support portal, a page builder, or a chatbot's shared-link feature.
- 3:26So spend an afternoon this week on it.
- 3:29Ask which of your own domains carry content you don't fully control.
- 3:33And ask whether your procurement checklists for AI tools ask vendors how they isolate user-generated pages fr
- 3:41om official downloads.
- 3:46Get software from an internal catalogue, or the vendor's documented address.
- 3:51Never from a paid search result, however correct the domain looks.
- 3:55That was Decrypted. The written version and every source are linked in the show notes.
- 4:00Thanks for listening.