Hoping to leave recent troubles behind, PaperCut has released another security maintenance update that replaces all previously published emergency patches. The patches were meant to definitively deal with a couple of serious security flaws – but they were still exploited in hundreds of AI-powered attacks.

PaperCut, a software system used by businesses, schools, and other organizations to manage printing, first disclosed the security defects, tracked as CVE-2026-82078 and CVE-2026-81578, as zero days on August 27th.

The vulnerabilities could allow remote unauthenticated attackers to bypass authentication and execute arbitrary code on vulnerable PaperCut NG and PaperCut MF instances.

They were patched the next day. However, PaperCut almost immediately urged users who had already installed the first emergency patch to do it again using an updated Emergency Patch (Release 2).

Now, the firm has published a new security maintenance release that replaces all previous patches.

“These are Regular Maintenance Releases (MR) that have gone through complete QA testing. They contain all of the security fixes issued in Emergency Patch Releases 1, 2, and 3, plus additional security hardening, and they have been through our standard release testing process,” the advisory says.

“These releases replace the emergency patches. If you are running an emergency patch build, move to a maintenance release. If you have not yet patched, upgrade now.”

The problem is that the 2 vulnerabilities were found to be quite easy to reproduce. Plus, multiple ways to bypass the initial patch were discovered – and cybercriminals seem to have done exactly that.

This week, threat intelligence firm GreyNoise revealed that a Russian-speaking threat actor has used AI to weaponize the flaws and break into at least 395 organizations in 48 countries, most of them concentrated in the US education sector.

According to GreyNoise, the attackers used hundreds of AI agents powered by OpenAI’s Codex harness and a DeepSeek model to target organizations at scale, while avoiding entities in Russia, China, Hong Kong, Thailand, Iran, and 23 other countries.

It is unclear if this actor is solely focused on access development to be handed off to other affiliated actors or if they will directly leverage their accesses to achieve follow-on objectives such as data theft or ransomware deployment,

GreyNoise said in a blog post.

To the researchers, this was another example that large language models are enabling cyber threat actors to move at greater speed and scale.

“The adversary went from an empty workspace to first achieving RCE against a real victim in just under 4 hours, first domain admin in an additional 2 hours, and once the full campaign launched, compromised at least 11 organizations in 26 seconds,” said GreyNoise.

In one instance, the attacker went from initial access to full domain administrator in just 7 minutes at a high school in the US.

And recently, another attacker used AI agents to breach an enterprise in less than 10 hours. Human operators would normally have taken around 2 weeks to complete the breach.

In August, another group of researchers said AI agents had carried out a 4-day cyberattack against Taiwanese government agencies with minimal human input. The agents autonomously mapped networks, found weaknesses, exploited systems, and collected data.