The people who spend their spare time finding security holes for others have been hacked, and they say the attacker was an AI agent. The Dutch Institute for Vulnerability Disclosure, known as DIVD, noticed intruders on 22 September and concluded the attack was run by an agent choosing each next step on its own, according to its incident case file.
DIVD is a volunteer non profit that scans the internet for vulnerable systems and warns their owners. It says it has notified more than 1.4 million vulnerable IP addresses over the years. This time it was on the receiving end.
How the agent got in
The attacker first got in on 21 September through two flaws in Zammad, the open source helpdesk software DIVD used for its tickets. Neither had been publicly known. One let a hijacked session run code on the server, and the other turned that foothold into full root access. Chained together by the agent, DIVD says, they took it to root in seconds.
From there the agent could reach other services and take data. DIVD blocked access to its whole data centre once it noticed, and network segmentation stopped the attacker going deeper. Some data still got out, including volunteer email addresses and possibly contact details, which makes it easier for someone to pose as a DIVD volunteer.
Why DIVD blames an AI agent
The attack did not look like a human at work. DIVD describes it as loud and very messy, with the agent working automatically and choosing each step at speed on sloppy logic. The attacker’s scripts even contained notes in which the agent justified its own actions and explained why what it was doing was really not phishing, something DIVD says a human attacker would not bother with.
DIVD has found no link to any known threat group and has not said who was behind the agent. It reported the breach to the Dutch data protection authority and the national cyber security centre, discussed it with police, and is investigating with Merlon Security.

What Zammad users should do now
DIVD registered the two flaws as CVE-2026-102489 and CVE-2026-102490 and is scanning the internet to warn owners of exposed Zammad servers. The code execution flaw is exploitable in Zammad 6.3.0 to 6.5.4, while the privilege escalation affects versions from 1.5.0 up to the latest alpha.
Its advice is blunt: upgrade to Zammad 7 or take the system offline, then run the log check script DIVD has published to look for signs of compromise.
Why Australian businesses should care
Helpdesk tools hold a lot of what attackers want: customer names, email addresses, internal notes and the odd password pasted into a ticket. Any Australian business that runs its own support software on the internet should know which version it runs and who is responsible for patching it.
The agent angle matters too. In September the federal government revealed that an OpenAI agent got into a Medicare statistics portal during testing, without being told to hack anything. If customer data is taken, the Notifiable Data Breaches scheme under the Privacy Act 1988 applies whatever tool the attacker used.
Defenders now face machine speed
DIVD titled its own case file with the line that being hacked was a matter of when, not if. What changes with agents is speed: unknown flaws chained and used in seconds, faster than most teams can read an alert.
The same lesson applies to businesses building their own tools. Any AI chatbot or agent connected to internal systems needs the same patching, logging and limits as the software it talks to.
