Fake AI Desktop Apps Abuse Signed OpenAI and Anthropic Binaries to Deliver Malware

| 2 Min Read
A campaign uses counterfeit AI app downloads and trusted online services to install malware, showing why software signatures alone can’t establish trust.

Signed software can still open the door to malware when attackers control what ships beside it. A campaign tracked as Yoba is using counterfeit ChatGPT and Claude desktop apps to exploit that gap, pairing genuine OpenAI or Anthropic executables with malicious files that run on Windows systems.

Check Point Research found fake GitHub repositories and websites advertising apps with names such as “GPT-5.6 Sol,” “GPT-6 Astra” and “Claude Opus.” The pages direct visitors to ZIP downloads hosted through GitHub releases. Despite lacking inflated stars or forks, some projects appeared near the top of Google results, which researchers attributed to search engine optimization poisoning.

The researchers reported malicious releases uploaded on August 30 and 31, and September 4. They described the sites as convincing, with search-targeted content and installer delivery tactics designed to make the downloads look credible.

Trusted executables, malicious DLLs

Each package places a real, digitally signed Setup.exe from OpenAI or Anthropic alongside a malicious file named winmm.dll. When Setup.exe runs, it loads that library, which Check Point calls YobaDropper. The signature validates the executable itself; it doesn't vouch for the other files in the package or everything the program loads.

Yash Kadakia, founder and CTO of Security Brigade, said the distinction matters: “A valid digital signature establishes the authenticity and integrity of a particular executable. It does not establish that everything the executable loads, or the package it was distributed with, is trustworthy.” He summed up the risk this way: “A signed executable is not the same thing as a trusted execution chain.”

YobaDropper delays activity for about 12 days unless a process named Setup.exe launches it. It also avoids systems with Russian, Belarusian, Kazakh and other CIS-region keyboard layouts, then requests administrator privileges. With elevated access, it copies a legitimate Windows Media Player binary into ProgramData, places another malicious DLL beside it and creates a scheduled task. The task claims Microsoft Corporation as its author and runs as Local System three minutes after startup.

Delivery split across familiar services

Once persistence is in place, a second component, YobaLoader, contacts attacker-controlled infrastructure for instructions and a payload. Check Point found a chain that uses a Telegram channel and an image to help produce a domain name, then queries Google’s DNS-over-HTTPS service for encrypted instructions in a DNS TXT record. The loader downloads its encrypted payload from Catbox, a file-hosting service.

In the activity examined, the payload was XMRig, a cryptocurrency miner that routes traffic through Tor and pauses while the user is active. Check Point said indicators suggest Russian origins but did not attribute the campaign to a named group. The loader can also fetch other Windows executables or PowerShell scripts, so cryptomining may not define the limits of future attacks. “Future waves may deliver different malware families,” the researchers wrote.

Using mainstream services makes individual connections harder to flag on reputation alone. Kadakia said Yoba spreads its communications across legitimate platforms instead of relying on dedicated command-and-control servers. That complicates detection for security teams focused on known malicious domains, IP reputation or conventional DNS monitoring.

The activity still leaves clues. Kadakia pointed to unexpected DLL loads by signed apps, scheduled tasks that launch copied system binaries, execution from ProgramData and unusual network traffic. Finding those signals requires teams to correlate endpoint and network data. Check Point also identified a Yoba variant from June that masqueraded as the Cursor AI coding assistant.

Make approved AI tools easier to get

For enterprise security teams, the campaign exposes a practical tension: employees may seek AI tools outside approved channels when internal access takes too long. “If employees can discover and install a useful AI tool in minutes but internal approval takes weeks, shadow IT becomes the predictable outcome,” Kadakia said.

He recommended reviewing permanent local administrator access on developer workstations, expanding application controls to cover DLL loading and execution behavior, and monitoring software from unverified repositories or user-writable folders. Developer machines deserve particular attention because they may contain source code, API keys, cloud credentials and routes into internal infrastructure.

“AI security cannot be limited to preventing employees from sharing sensitive information with AI models,” Kadakia said. The Yoba campaign points to a wider operational need: manage how AI software is sourced, installed and allowed to run across company systems. OpenAI and Anthropic did not immediately respond to requests for comment.

Source: Christopher Rodriguez · www.csoonline.com

Comments

Please sign in to comment.