AI Agent Chains Zammad Zero-Days to Breach DIVD
Attackers used two Zammad helpdesk zero-days and an AI agent to go from user to root in seconds at the Dutch Institute for Vulnerability Disclosure.
The Dutch Institute for Vulnerability Disclosure (DIVD) — a volunteer group that finds security holes in other people's software — has revealed it was breached through two zero-days in its own helpdesk system, Security Affairs reports.
What happened
Attackers got in through Zammad, an open-source ticketing platform. With Merlon Security, DIVD found two previously unknown flaws:
- CVE-2026-102489 — a session-hijack vulnerability leading to remote code execution as the Zammad user, affecting Zammad versions 6.3.0 to 6.5.4, with a CVSS score of 9.4 (SecurityWeek).
- CVE-2026-102490 — a local privilege-escalation vulnerability allowing the Zammad user to escalate to root, affecting all versions including the latest alpha, also scored 9.4 (SecurityWeek).
Chained together, they let attackers hijack sessions, run code remotely and go from an ordinary Zammad account to root access in seconds. From there, they reached other services and stole data.
Timeline of the breach
DIVD's published case files give a detailed timeline of the incident:
- 21 September 2026: First access by the malicious actor on DIVD's systems.
- 22 September 2026: DIVD becomes aware of the malicious activity and blocks access to all systems in its datacenter; an incident response team is formed and a forensic investigation begins together with Merlon Security.
- 24 September 2026: DIVD publishes its first public statement, saying it has informed directly involved parties, reported the incident to the Dutch data protection authority (Autoriteit Persoonsgegevens) and the National Cyber Security Centre, and discussed options with police.
- 30 September 2026: DIVD publishes a detailed casefile (DIVD-2026-00014 and DIVD-2026-00015) naming the two zero-day CVEs and explaining how they were chained.
In its own words, posted on LinkedIn on 24 September, DIVD said: "Security people always say it's not a matter of if, but when you get hacked. It took us (almost) seven years but we can now say that we're the hackers that got hacked" (DIVD CSIRT).
The irony is hard to miss. DIVD's entire mission is scanning the internet for vulnerable systems and quietly warning their owners before criminals find them. Its own helpdesk — the system it uses to coordinate those warnings — turned out to be the vulnerable system.
The AI factor
DIVD said the speed of the attack was down to the use of an AI agent. Steps that would take a human attacker minutes or hours — testing, chaining and escalating — happened almost instantly. DIVD described the attack as "loud and very, very messy," driven by an AI agent that moved autonomously and decided its next steps without external intervention or direction, according to BleepingComputer's reporting on DIVD's follow-up disclosure (BleepingComputer).
Notably, DIVD says it was able to retrieve extensive details about the attack precisely because the AI agent left behind clear explanations of its own decisions in logs, which allowed investigators to reconstruct much of the incident after the fact — an unusual silver lining that let defenders understand an AI-driven attack chain in detail rather than being left to guess.
DIVD summarized the technical chain in its own words: "Used together, they allowed the attackers to hijack sessions, run code remotely, and escalate privileges from the Zammad user to root, in seconds, due to the agentic part of this hack" (SecurityWeek).
This is the part of the incident that should worry defenders most. Security teams have long relied on the assumption that exploiting a fresh vulnerability takes an attacker time: time to understand the flaw, time to write the exploit, time to move through the network. An AI agent compresses all of that into the gap between two log lines. By the time an alert fires, the attacker may already be done.
What limited the damage
Despite the speed of the attack, DIVD says it was largely contained. "Due to network segmentation and incident response actions, the threat actor did not move deeper into the network," BleepingComputer reported, and DIVD itself noted: "We've found signs of compromise that we're still looking into, and, until we can prove otherwise, we assume breach. Still, stopping the attackers is a win, and in a situation like this you take every win you can get" (SecurityWeek). This illustrates that even against an AI-accelerated attacker, fundamental defensive practices — in this case, network segmentation that stopped the intruder from pivoting further into DIVD's infrastructure — still provided a meaningful backstop.
Background on Zammad and DIVD
Zammad is an open-source, AI-powered helpdesk and customer-support ticketing platform, self-hostable or available as a managed service, which the vendor says is used by over 2,000 customers and 55,000 users, including organizations such as De'Longhi, Amnesty International and Nextcloud (BleepingComputer). DIVD itself is a Dutch nonprofit, run largely by volunteer security researchers, that has operated for nearly seven years performing coordinated vulnerability disclosure — meaning it proactively scans for and reports vulnerable, internet-exposed systems to their owners before criminals can exploit them.
Why it matters
If an organization of professional vulnerability hunters can be breached this way, anyone can. AI-driven attacks shrink the time defenders have to spot and stop an intrusion from hours to seconds, which breaks the detection-and-response model most security teams are built around.
There is also a supply-chain angle. DIVD holds sensitive information about unpatched vulnerabilities in thousands of organizations. A breach of its systems is not just one organization's problem — it potentially hands attackers a map of everyone else's open wounds. DIVD has not suggested its vulnerability reports were the target, but the possibility alone underlines how much trust is concentrated in disclosure organizations.
What Zammad users should do
DIVD urges everyone running any version of Zammad to:
- Update to version 7 immediately, or take the system offline.
- Review logs for unusual sessions and new admin accounts.
- Rotate passwords and API keys stored in or connected to Zammad.
- Download and run DIVD's published verification script against Zammad log files to check for indicators of compromise (DIVD CSIRT).
Helpdesk systems deserve special attention beyond this one incident. They are internet-facing by design, hold credentials and sensitive conversations, and are often connected to email and identity systems — a near-perfect pivot point for an attacker.
What readers should do
- If your organization runs Zammad in any version, treat patching to version 7 (or taking it offline) as urgent, not routine, since both CVEs are already known and have been exploited.
- Apply network segmentation around helpdesk and ticketing software generally, so that even if one of these systems is compromised, attackers cannot easily move into more sensitive internal systems — this is specifically what DIVD credits with limiting the damage in its own incident.
- IT and security leaders should start treating "agentic AI" attacks as a realistic category of threat in incident-response planning, since the speed of automated exploitation may leave far less time to detect and respond than traditional human-driven intrusions.
- If you use any disclosure or bug-bounty coordination service, consider asking the provider directly what safeguards protect the sensitive vulnerability data they hold, given the supply-chain risk this incident highlights.
Sources & Verification
Additional reporting
- Security Affairssecurityaffairs.com
- DIVD CSIRT – DIVD-2026-00014csirt.divd.nl
- DIVD CSIRT – DIVD-2026-00015csirt.divd.nl
- DIVD CSIRT – "It was a matter of when, not if"csirt.divd.nl
- BleepingComputerbleepingcomputer.com
- SecurityWeeksecurityweek.com



