Hackers Are Using AI to Attack Windows and Linux Servers at Scale

• The Mess: A Chinese-speaking cybercrime group tracked as UAT-10147 is using agentic AI to automate attacks against exposed Windows and Linux web servers. Cisco Talos found evidence of AI-assisted exploit refinement, reconnaissance, payload generation, troubleshooting and post-compromise automation — with a target list containing roughly 170,000 URLs.

This is not “hackers asked ChatGPT how to write a PowerShell script.”

It’s considerably more interesting.

UAT-10147 is using AI as part of the operational workflow.

An exploit fails?

The tooling can refine it.

Something doesn’t work?

The attacker can generate troubleshooting logic.

Access is obtained?

Automation takes over reconnaissance, persistence and payload deployment.

And the targets aren’t theoretical.

Talos identified compromised servers belonging to organizations in government, education, media, technology and gaming, across multiple countries. The group’s infrastructure contained a list of around 170,000 URLs, divided into smaller batches to make scanning and exploitation more efficient.

The attackers aren’t waiting for some mythical zero-day.

They’re exploiting known vulnerabilities on Internet-facing servers.

That’s the important bit.

Old bugs.

New automation.

Much faster attacks.

• The Damage: AI-assisted automation dramatically increases the number of vulnerable servers an attacker can probe, exploit and maintain access to, turning an unpatched Windows or Linux server into a potential entry point for malware, data theft, SEO fraud and further network compromise.

Talos observed different attack chains depending on the platform.

On Windows/IIS, attackers used RCE to deploy scripts, escalate privileges, modify Microsoft Defender exclusions and establish persistence. They also deployed malware such as QuasarRAT and BadIIS, with additional backdoors including SPECTRE.

On Linux, the attackers used RCE to get the initial foothold, deployed web shells and then used known local privilege-escalation exploits to move toward higher privileges.

That’s where this gets nasty.

The AI isn’t replacing the exploit.

It’s making the whole attack process cheaper and more scalable.

Think about the economics.

A human attacker investigates one server.

An automated system can work through thousands.

One failed exploit isn’t necessarily the end.

The system can try another approach.

And another.

And another.

The attacker effectively turns vulnerability management into a numbers game.

If 170,000 targets are available and only a tiny percentage are exploitable, that’s still potentially a lot of compromised machines.

The campaign also demonstrates something defenders should take seriously: AI is being used after the initial compromise, not just before it.

Talos describes iterative exploit refinement, adaptive troubleshooting, exploit validation and post-exploitation automation — a move toward semi-autonomous offensive operations rather than simple AI-assisted scripting.

• The Fix: Patch Internet-facing Windows and Linux applications aggressively, remove unnecessary public exposure, monitor for unexpected web shells and privileged-account creation, and treat unexplained Defender exclusions, scheduled tasks and unusual server-side process activity as potential compromise indicators.

The lesson isn’t “install an AI firewall.”

It’s much less glamorous.

Patch the damn server.

The attackers are specifically taking advantage of publicly disclosed vulnerabilities.

That means organizations already have an advantage: the bugs are known.

Use it.

For Windows/IIS environments, investigate unexpected changes to Defender exclusions, IIS configuration, scheduled tasks and local administrator/RDP accounts. Talos observed those exact behaviors during the campaign.

For Linux servers, look for unexpected web shells, new privileged accounts, suspicious kernel modules and unusual processes spawned by web applications.

And don’t forget the web layer.

A perfectly patched operating system doesn’t save you if the application running on top of it is vulnerable.

Bugstoday Opinion

This is one of the more important security stories this week because the vulnerability itself isn’t the headline.

The headline is the economics.

Attackers are taking the same old vulnerabilities we’ve been warning about for years and adding automation.

That changes the equation.

A bug that was previously “probably won’t be exploited against my obscure server” becomes much more interesting when an automated system can scan enormous target lists and try exploitation without getting tired, bored or distracted.

AI doesn’t need to invent new exploits to make things worse. It just needs to make existing attacks cheaper.

And that’s exactly what UAT-10147 is demonstrating.

Bugstoday verdict: the AI isn’t the scary part. The scary part is 170,000 Internet-facing targets, known vulnerabilities and attackers automating the boring work. If your server is exposed and unpatched, you’re basically volunteering for the experiment.