# Yoba uses ChatGPT and Claude signatures to slip AI malware through

> Check Point uncovers Yoba: fake repos with supposed ChatGPT and Claude apps that abuse signed binaries to install a miner.
> By José Maldonado · Cybersecurity · 2026-10-29

Someone searches for "ChatGPT for desktop" on Google, finds a GitHub repository with that name, downloads the ZIP and runs the installer. The installer has a digital signature from OpenAI. And the package, nevertheless, installs a cryptocurrency miner. That sequence is what the investigation published by [Check Point](https://blog.checkpoint.com/research/yoba-malware-abuses-signed-chatgpt-and-claude-binaries-to-target-ai-users/) describes about Yoba, a campaign that relies on three things people already consider normal: GitHub, Google search and a valid signature.

The important part of the finding is not the miner. The researchers say it clearly: the final payload observed was a Monero miner in its 6.26.0 version, but YobaLoader is a generic loader capable of delivering other Windows executables or PowerShell scripts. The interest lies in the infection chain, which turns the search for AI tools into the entry vector.

Nor is there a breach at OpenAI or Anthropic. Neither company has been compromised, and the legitimate binary that runs is authentic. What fails is the trust the user places in an entire package because one of its pieces is signed. That nuance separates a third-party incident from a provider breach, and explains why the campaign works without touching any third-party server.

## Repositories with names that sound like real products

The attackers created repositories and GitHub Pages sites advertising supposed desktop versions of AI products, with names like "GPT-5.6 Sol", "GPT-6 Astra" and "Claude Opus". The download links pointed to ZIP files hosted as releases inside GitHub.

The repositories showed no stars or inflated forks, the classic trick to look popular. Even so, they appeared prominently on Google, and Check Point assesses the pattern as search result poisoning: content built to appear before anyone actively searching for AI software.

That tactic shifts the attack to the moment of decision. The user already wants to install something, is already searching for a name and a link, and the attacker only has to be in the right place. No one has to be convinced to open a strange file; it is enough to offer the file they were looking for.

## The valid signature and the DLL that is not

The package mixes malicious components with legitimate software. The ZIPs imitating ChatGPT include an authentic \`Setup.exe\` signed by OpenAI; those imitating Claude include a signed Claude executable. When launched, those binaries load a malicious \`winmm.dll\` placed alongside them using the DLL side-loading technique.

The result is a chain with two uncomfortable truths. The first: the signature tells the truth about the executable and nothing about what surrounds it. The second: the side-loading technique still works in 2026 as effectively as a decade ago, because the operating system still trusts the program's folder.

That pattern matches other recent incidents. The [Siemens Software Center CVE](https://observatorioblockchain.com/en/cybersecurity/siemens-software-center-cve-2026-110936-clickfix/) exploited with ClickFix used the same idea: legitimate binary, third-party library and execution with the trust of the former. The vendor and the sector change, not the logic.

## A loader that hides from analysts

YobaDropper, the first stage, incorporates several checks before doing anything. It verifies that it was launched by the expected \`Setup.exe\`, introduces a long wait if the check fails, applies regional language controls and requires administrator privileges to continue.

Those conditions are designed to complicate automated analysis. A system that runs samples at full speed will not have the expected environment and the malware will stay still, which delays detection and makes many samples look harmless.

When the checks pass, the dropper deploys YobaLoader alongside a legitimate Windows executable and creates a scheduled task so that this executable loads the malicious library after boot. Persistence does not depend on its own service or a conspicuous registry entry, but on a known process that starts on its own.

## Telegram, DNS and Catbox as infrastructure

The most original part of the analysis is the communication. YobaLoader does not connect to a dedicated command and control server. Instead, it queries a Telegram channel and downloads the profile picture. The data from that image is used to generate a domain, so the attacker changes the domain by changing the channel photo.

It then queries Google's DNS over HTTPS for a TXT record of that domain. It validates and decrypts the response and, with it, retrieves an encrypted payload from an external host such as Catbox.

Each piece is a legitimate service. Blocking Telegram, Google DNS or a file hosting service carries an enormous operational cost for any company, and the attackers know it. By mixing commonly used services with dynamically generated domains, the malicious infrastructure becomes hard to distinguish from normal traffic.

Stage

What it does

Service used

Discovery

Repos with supposed AI apps and links to ZIP

GitHub and GitHub Pages

Delivery

Legitimate signed binary plus malicious \`winmm.dll\`

OpenAI, Anthropic (no breach)

Persistence

Scheduled task that loads the DLL alongside a system executable

Windows

Direction

Profile image to generate the domain

Telegram

Resolution

TXT record of the generated domain

Google DNS over HTTPS

Payload

Encrypted download of the next module

Catbox

## The miner and what comes next

In the campaign analyzed, the final payload was a Monero 6.26.0 miner that can communicate over Tor and that pauses mining when it detects that the user is working, so as not to give away its presence through consumption.

That behavior says a lot about the objective: silent monetization. A miner that slows down the machine is discovered in a day; one that steps aside when there is activity can go months without being noticed.

The researchers insist on not stopping at that label. YobaLoader can deliver other executables or scripts, so treating the incident as a miner case is a scope error. An infected machine has a loader with persistence and connectivity, and whatever it has brought is the least important piece.

## What to do in an organization?

The first measure is procedural: install AI software only from the provider's channels. A signed binary inside a package downloaded from a third-party repository proves nothing about the rest of the package.

The second is about permissions. If an employee needs administrator privileges to install their favorite tool, the problem is not the tool, it is the granting model. Limiting those permissions reduces the surface where a dropper can operate.

The third is about detection. Looking only for the malicious file is no use here, because the chain goes through legitimate processes. What needs to be monitored is what a signed executable loads and what scheduled tasks are created after an installation.

And the fourth is about inventory. Knowing which AI applications run on the machines, and where they came from, is today part of security and not a taste for order. If the organization cannot answer that question, the Yoba campaign will find a gap without needing to exploit any vulnerability.
