Until recently, a serious cyberattack needed a skilled person, time and patience. Last week, a security firm documented an attack in which a single attacker put hundreds of AI agents to work and broke into 395 organizations across 48 countries, many of them within seconds. This is not news for specialists only: the victims include shops, professional services offices and coworking spaces, exactly the kind of small or medium business you run.
What happened, in short
On August 27, 2026, PaperCut, the maker of a print management program, published an urgent warning and released the first security patch for two serious flaws (PaperCut). Four days later, on August 31, an attacker used AI agents to build, test and launch the attack at scale (GreyNoise).
The result: at least 440 compromised servers at 395 organizations in 48 countries (GreyNoise). The attacker stole passwords from 280 victims and gained full control over the internal network at 12 of them (Help Net Security).
The tools were nothing exotic: a commercial coding agent from OpenAI and a DeepSeek model, combined with publicly available security testing programs (Help Net Security). In other words, the same kind of AI tools we write about here as a way to automate your quotes or customer replies.
Why this concerns you, even if you have never heard of PaperCut
Three things have changed compared with previous years.
Speed. At the peak of the campaign, the agents compromised 11 organizations in 26 seconds, and at a US high school they went from first entry to full control in seven minutes (GreyNoise). There is no more "we will deal with it next month".
The time window. Only a few days passed between the patch and the attack. If your company updates its software "when someone gets around to it", you are exposed precisely in the period when attackers are most active.
The target. A human attacker picks large victims, where the effort pays off. An AI agent does not get tired and does not select: it hits everything it finds open on the internet. The victims included 38 companies in retail and professional services and 29 in real estate, coworking and hospitality (GreyNoise). Small businesses are no longer "too small to matter".
One detail says a lot: the attacker had configured a list of countries the agents were supposed to avoid, yet victims appeared in those countries anyway (GreyNoise). Not even the person who launched them could fully control them.
What you can do in your company this week
You do not need an IT department. You need a few clear habits, building on the basics we covered in NIS2 in 2026: security as a business requirement.
Make a list of what is exposed to the internet. Print server, a NAS with files, security cameras, a router with remote access, a management program left open "so we can work from home". Ask whoever handles your IT: which of our services can be reached directly from the internet? If the answer is "I don't know", that is the first thing to fix.
Close what you do not use. A service that is not reachable from outside cannot be attacked from outside. For what really has to stay reachable, put it behind a VPN or, at the very least, restrict it to known addresses.
Update within the same week, not the same quarter. When a vendor releases an "urgent patch", it is a signal that the flaw is already being used or soon will be. Set a simple rule: security updates are applied within 72 hours at most. In the PaperCut case, companies that followed such a rule had time to protect themselves before August 31 (GreyNoise).
Two-step login everywhere. The attackers stole passwords from 280 organizations (Help Net Security). A stolen password without a second factor (a code on your phone, an authenticator app, a physical key) is an open door. With a second factor, it is just a stolen password.
A backup you have actually tested. Not "we have backups", but "last month we restored a file from backup and it worked". Keep one copy separate from the company network, so it cannot be wiped along with everything else.
One responsible person, with a name. Even if IT is outsourced, someone inside the company needs to know who receives security alerts from vendors and who decides that a patch gets applied. Without a name, urgent messages stay unread.
Careful with your data: the same AI, two faces
The story above shows that AI tools are powerful no matter who uses them. When you use them in your company, think of them as a very capable external contractor, but one you do not hand everything to.
Do not send an AI assistant your customers' personal data (names together with ID numbers, addresses, phone numbers), contracts with confidential clauses, passwords, access keys or bank statements. If you need help with a contract, replace names with "Client A" and amounts with approximate values. The result is just as good and the risk is much lower.
For a company in Romania, GDPR is clear that you remain responsible for customer data even when you hand it to a vendor. Check where the data is processed (ideally in the EU), whether the vendor uses it to train models and whether you have a signed data processing agreement. We detailed the checklist in GDPR and AI: what data you can and cannot give an AI assistant. These rules are not bureaucracy: they are exactly the line of defense that limits the damage if an account or a vendor gets compromised.
And, mirroring the attack above: if you put AI agents to work in your company (invoicing, emails, reports), give them only the access they strictly need and keep a human in the loop for any irreversible action (we wrote five tips on keeping control of an AI agent). The attacker in this story lost control of his own agents. You do not want the same to happen with yours.
Conclusion and one concrete step
Automated AI attacks are no longer a future scenario, they started in August. The good news is that basic defenses, applied on time, still work: even in this campaign, a properly configured web application firewall stopped an attack attempt (GreyNoise).
Today's step: send a message to whoever handles your company's IT with two questions. Which of our services are reachable directly from the internet? When were the latest security updates applied to them? The answer tells you in five minutes how exposed you are.
If you want a structured check, at Marincea Digital Consulting we run IT audits and help you set simple rules that your company keeps following after we leave.
Read next
NIS2 in 2026: security as a business requirement, not just IT
GDPR and AI: what data you can and cannot give an AI assistant
Put an AI agent to work in your firm without losing control: 5 tips