Accessibility links Skip to main content

Ethical hackers got hacked themselves, with AI: “The fingerprints were everywhere”

A Dutch organization that hunts for vulnerabilities in software was hacked itself. It turned out to be an AI-driven attack, something the ethical hackers at DIVD had never experienced before. Volunteer Frank Breedijk offers a look behind the scenes. “It’s like a busload of hyperintelligent toddlers storming in.”

Frank Breedijk is getting ready for an evening of Netflix on the couch that Tuesday night in late September. But a phone call changes his plans: “We need you as crisis manager.” On the other end of the line is DIVD, the organization of ethical hackers where he volunteers. The organization has been searching for vulnerabilities in software for years. After the call, technical investigation follows, and Breedijk concludes: “We’ve been hacked.”

As a cyber expert, Breedijk has seen a lot. He has more than two decades of experience and works as chief information security officer (CISO) at Schuberg Philis, a large Dutch IT company. He had expected that DIVD would be hacked at some point. But this was no ordinary intruder. It turned out to be AI agents that, among other things, got hold of the ethical hackers’ email addresses. “It was different from a human hacking. It was much harder to fight.”

Breedijk: “I don’t like spreading fear around cybersecurity. But this kind of attack with AI will only get worse over the coming months.” That’s why DIVD, which stands for Dutch Institute for Vulnerability Disclosure, will soon publish an extensive report on this AI attack. The organization, which has about 150 volunteers, wants to use it to help companies and institutions.

The risks of cyberattacks using AI have grown this year. More advanced AI models have no trouble finding vulnerabilities in software. With AI agents, pieces of software that can carry out tasks automatically, attacks now also move much faster. Cyber defenders have less time to respond. Warnings about these risks have been piling up in recent months.

After the call to Breedijk, DIVD activated its cyberattack protocol for the first time in its seven-year existence: contain the situation, eliminate the attackers, and then restore everything. The attacker was isolated in one part of the network. The breach was in the open-source software Zammad, which handles email for registering new vulnerabilities and for questions about vulnerable systems and leaked passwords. A notification was filed with the Dutch Data Protection Authority, and a report was made to the police.

“A human wouldn’t do that”

Everything went by the book, but then something stood out. “We saw many more traces than normal. A human hacker would leave traces in thirty files at most in that time. We found them in more than three thousand. The fingerprints were everywhere.” These are pieces of programming code (scripts) that attackers plant in systems to carry out a task. That also explained why the attacker was able to do so much at once.

“Those scripts turned out to include explanations of how they work. Somewhere it said: ‘this should only run during office hours in the Dutch time zone.’ And then more comments on that behavior. Everything was neatly documented. A human wouldn’t do that. But AI language models have learned to program this way.”

Furthermore, the attackers had broken in via a complex chain involving two zero days in Zammad, unknown vulnerabilities for which no fix exists yet. “The fact that they had two of them and were willing to use them was striking. Previously, each was worth about €1 million to cybercriminals. They would have saved them for a more attractive target.” It shows how “zero days” are subject to inflation because of AI.

Once DIVD had isolated the attacker, the attacker changed strategy. The hacker tried to get further by breaking into the companies that keep DIVD up and running, the hosting providers. Through the Zammad vulnerability, they gained access to the mailbox for the vulnerability registration system. They could watch along, without seeing all the details. At the other hosting provider, an outage occurred that also affected other customers.

“The attackers could have done much more at one of the hosting companies. A human would most likely have shifted their attention. But they kept focusing on us.” That was the last piece of the puzzle. An AI agent sticks to its assignment and carries it out rigorously. That became clear this summer, when AI agents at OpenAI and Anthropic got out of their test environments and started hacking companies and governments to achieve their goal.

The conclusion was inescapable: AI agents. But who was directing them? “AI doesn’t attack on its own; there’s always a human behind it giving the assignment,” says Breedijk. “But finding out who is the hardest part of any cyberattack.” A number of scenarios could be ruled out, such as an escaped AI agent from a company or an employee who wanted to test something and it went wrong. “It was probably someone who wanted insight into the vulnerabilities we find.”

The attacker used Artex, a platform that can send swarms of AI agents out to hack. It doesn’t even require the most advanced AI model, says Breedijk. The agents look for weaknesses in security and exploit them. Artex is freely available (open source) and has Chinese roots. Last month, Microsoft warned about hacks using Artex at unnamed cloud services. The attackers were linked to the Chinese state and gained full access in less than three hours, according to Microsoft.

Cookie jar

Artex’s AI agents work noisily, according to Breedijk. “Artex is like a busload of hyperintelligent toddlers descending on your house to find the cookie jar.” Most comparable platforms work more subtly. But what all these AI agents have in common is that they tirelessly try to achieve their goal. “And that they’re inscrutable. With a human, you can predict what they’ll do. But when Artex sends those toddlers after you, they might also, quite illogically, look for the cookies in the cabinet with the cleaning supplies.”

The Artex attacker viewed the DIVD mailbox nine times, scanned the internal network, pulled login information from a database, and tried to change passwords, including for DIVD accounts at the hosting company. All of that at the same time. Data was stolen: employees’ email addresses and possibly also information about reported vulnerabilities. But Breedijk can’t share all the details yet.

The report on the hack will be “not always fun” for his organization. During the investigation, it turned out that even DIVD had missed a “patch,” a security update for software. The first response to the attack also took nineteen hours. “As a volunteer organization, we didn’t have enough people during office hours to respond quickly. The attacker had also blocked access to parts of the network.”

Breedijk expects the turbulence around AI attacks to subside. “Cyber defenders and companies now have to adapt to a new reality. We’ve never experienced this before. But I’m confident we’ll manage. We’ll get used to the pace and intensity of the attacks. The use of AI will also help with defense.”

Pulling all-nighters

The Hague-based cybersecurity company Eye Security recently assisted a customer (“a large European e-commerce company”) that was also attacked with AI agents via Artex. Because of “the scale of the attack,” the cyber defenders deployed AI agents themselves, with restrictions and under supervision. For a while the attacker seemed to back off, but then it started again. The tokens used to purchase AI computing power had probably run out for a while, says Eye Security. The company is holding a webinar about this attack today.

What he went through was no small thing for Breedijk. “Someone breaks into your systems. You have a crisis to handle and a volunteer organization to hold together. Quite a few people pulled all-nighters. Everything is very impactful.”

Source: FD. Jasper Houtman is technology editor at the FD.