Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentAI-enabled attacks

DIVD: an agentic AI attack chains two Zammad zero-days in seconds

CriticalDIVDZammad

On September 21, 2026 attackers broke into the Dutch Institute for Vulnerability Disclosure through its Zammad help desk, chaining an unauthenticated remote code execution flaw with a privilege escalation to root. DIVD says the modus operandi was one it had not seen before and points to an agentic AI-powered attack: after every action the agent decided the next step itself. Data was taken after a pivot to other services; network segmentation limited the damage. DIVD published details on October 1. Both Zammad flaws were added to the Known Exploited Vulnerabilities catalog on October 2, and on October 5 Zammad recommended upgrading to 7.2.0 and limiting server access to trusted administrators.

Replay the attack step by step in the Learning Center

How it happened

  1. CVE-2026-102489 let an unauthenticated attacker hijack sessions and run code on the Zammad server.
  2. CVE-2026-102490 escalated that foothold to root; DIVD says the chain ran in seconds because of the agentic part of the attack.
  3. An autonomous agent then handled post-exploitation, loudly and messily, pivoting to other services and exfiltrating data.

Impact

A vulnerability-disclosure organization was breached through a support tool many companies also expose; DIVD is treating its network as compromised until proven otherwise.

The fix

Zammad 7 is not exploitable in practice; DIVD's advice is to upgrade to 7 or take Zammad 6.x off the internet. Both CVEs were zero-days at the time of the attack.

Lessons

  • Help desk and ticketing systems are internet-facing by design and now get chained by agents faster than an advisory can be read.
  • Segmentation turned a full breach into a contained one; it is the control that still works when the perimeter falls in seconds.
  • Assume breach after an agentic intrusion: the agent tried many paths, and logs show only the ones it took.

Sources