An AI Agent Broke Into the Organization That Warns Everyone About Vulnerabilities

On September 21, 2026, the Dutch Institute for Vulnerability Disclosure (DIVD) was breached through two unknown flaws in Zammad, the open-source helpdesk it uses to handle reports. The attacker went from a hijacked session to root in seconds. DIVD says it was an autonomous AI agent: fast, loud, messy, and still successful. Here is what happened, what is still unknown, and what every security team should change.

For information only. This article is shared to inform defenders about a real intrusion and the vulnerabilities behind it. It is based on DIVD's public statements as reported by BleepingComputer and Help Net Security, and on the Cloud Security Alliance research note of October 3, 2026. It contains no exploitation details.

The Watchdog Got Bitten

DIVD is a Dutch non-profit, staffed mostly by volunteer security researchers, that scans the internet for vulnerable systems and warns their owners. Its whole purpose is to get to a flaw before attackers do. On September 21, 2026, an attacker got to DIVD first.

The entry point was Zammad, the open-source helpdesk and ticketing platform DIVD uses. The attacker chained two flaws nobody knew about, CVE-2026-102489 and CVE-2026-102490, to hijack a session, run code on the server and escalate to root. According to DIVD, that took seconds. From there, the attacker reached other services and read and exfiltrated data.

Then came the part that turned a breach into a headline. DIVD said the operation looked like the work of an autonomous AI agent, one that chose each next step itself after every action. In DIVD's words, the attack was "loud and very, very messy."

2 Zammad zero-days chained from session hijack to root
Seconds Time from initial session hijack to root, per DIVD
9.4 Chain severity assigned by Sysdig (8.7 and 8.5 individually)
v7 Zammad version DIVD recommends, or take the instance offline

How the Attack Unfolded

Putting together DIVD's updates and the CSA research note, the timeline looks like this:

DateWhat happenedStatus
Sept 21The attacker exploits the Zammad chain, reaches root and accesses other servicesBreach
Sept 22-23DIVD reproduces and analyzes the flaws, with incident response firm Merlon SecurityInvestigation
Sept 24DIVD discloses the hack, notifies Zammad and reports to the Dutch Data Protection Authority and NCSC-NLDisclosure
Sept 26Scanning for exposed instances and notification of other potential victims beginsWarning others
Sept 30DIVD publicly ties the breach to the two Zammad zero-days and an AI agentPublic details
Oct 8Zammad advises upgrading to version 7.0 or later and says it is working on a fix for the escalation flawFix in progress

Network segmentation and DIVD's incident response kept the attacker from moving deeper. The CSA note reports that confirmed exposure includes volunteers' DIVD email addresses and possibly contact details, which creates a risk of impersonation. DIVD says it assumes it is still breached until it can prove otherwise.

The Two Zero-Days, and Who Is Exposed

CVE-2026-102489 is a session hijack that leads to remote code execution as the zammad service user. It affects Zammad 6.3.0 through 6.5.4. Versions 7.0.0 to 7.1.3 contain the flaw, but DIVD says environmental conditions prevent exploitation there.

CVE-2026-102490 is a local privilege escalation that lets the zammad user become root. It is reported to affect versions from 1.5.0 through 7.1.0-alpha. Zammad says it cannot be exploited remotely on its own, and links it to a confirmed issue in the packager.io packaging service.

On their own, each flaw is serious. Chained, they turn an internet-facing helpdesk into a root shell. Researchers at Horizon3 have since published a technical breakdown and proof-of-concept for CVE-2026-102489, which means the window for unpatched instances is closing fast.

01
The Target Choice

Your Helpdesk Is a Crown Jewel

Ticketing systems rarely get the same attention as identity providers or production databases. Yet they hold credentials pasted by users, email integrations, API tokens, customer data and, in DIVD's case, vulnerability reports about other organizations. They are also internet-facing by design. An attacker that owns the helpdesk owns a map of everything the organization is worried about.

Why DIVD Believes It Was an AI Agent

DIVD has not named an actor or a model. Its conclusion comes from how the attacker behaved:

That last point is the irony of the case. The behavior that exposed the agent is exactly the behavior a better agent will not have. Defenders rebuilt the chain because the attacker narrated it.

What to read with care. The AI-agent attribution is DIVD's own assessment based on observed behavior, not a confirmed attribution. No actor or model has been identified. It is not known whether the agent discovered the zero-days itself or only used them, how initial access was obtained, or whether this was part of a larger campaign or a capability test. The CVE identifiers and CVSS scores could not be found in the official CVE record or Zammad's advisory archive when the CSA note was published, and some outlets reported different scores. The full scope of the data taken is still being assessed.

Part of a Pattern

The CSA calls this the second publicly documented autonomous AI-agent breach in under three months, after an AI model broke into Hugging Face in July. It also fits what this blog has tracked all year: an agent swarm breaching hundreds of organizations through PaperCut, frontier agents building their own attack chains, and AI closing the gap between lone hackers and nation-state teams.

The common thread is not sophistication. The DIVD agent was clumsy. What changed is that clumsy is now enough, because the agent tries everything, immediately, and only needs one chain to work.

What I Would Do Now

1. Find Every Zammad Instance

Include internal, test and forgotten deployments. Upgrade to version 7 or later, or take vulnerable instances offline. Before updating, copy application and network logs, as NCSC-NL advises, and run DIVD's published log-check script to look for indicators of compromise.

2. Rotate What the Helpdesk Could Reach

If you ran an affected version, assume its secrets are burned: API tokens, email integration credentials, database passwords. Rotate them, and warn staff that leaked contact details can fuel convincing phishing.

3. Segment Support Tools From Core Systems

Segmentation is the control that saved DIVD. Put the helpdesk in its own zone, apply default-deny egress, and make sure a root shell on that server does not mean a path to everything else.

4. Detect Behavior, Not CVEs

Zero-days have no signature. Alert on what the chain looks like instead: an application process spawning a shell, a service account becoming root, sudden outbound connections, and bursts of actions in seconds. Machine speed is itself a signal.

5. Pre-Authorize Automated Containment

When an attack finishes in seconds, waiting for a human to approve isolation is too slow. For high-confidence alerts, such as a service account gaining root, let your tooling isolate the host first and ask questions second. Measure time to contain, not only time to detect.

6. Plan for Quieter Agents

Do not build your detection around this agent's mistakes. Assume the next one will not leave comments, will not collide with itself, and will pace its actions to blend in. Test your defenses against that version.

An attacker no longer has to be good to win. It only has to be fast and persistent. The defenses that held at DIVD were architectural: segmentation and a prepared response team. Those work whether the attacker is a human, an agent, or a swarm.

Frequently Asked Questions

What happened to DIVD?
On September 21, 2026, an attacker chained two zero-day vulnerabilities in Zammad, the helpdesk DIVD uses, to hijack a session, execute code and escalate to root within seconds. It then accessed other services and exfiltrated data before network segmentation and incident response contained it.
Was it really an AI agent?
That is DIVD's assessment, based on the attacker's speed, self-directed decision-making, sloppy logic and verbose explanatory comments in its scripts. No actor or AI model has been publicly identified, and the attribution is not confirmed.
Which Zammad versions are affected?
CVE-2026-102489 affects versions 6.3.0 through 6.5.4; versions 7.0.0 to 7.1.3 contain it but are reported as not exploitable due to environmental conditions. CVE-2026-102490 is reported to affect versions 1.5.0 through 7.1.0-alpha. Zammad advises upgrading to version 7.0 or later.
What data was exposed?
Confirmed exposure includes volunteers' DIVD email addresses and possibly contact details, according to the CSA research note. DIVD is still assessing the full scope and says it assumes breach until proven otherwise.
I run Zammad. What should I do first?
Copy your application and network logs, run DIVD's log-check script for indicators of compromise, then upgrade to version 7 or later or take the instance offline. If you ran an affected version, rotate every credential the helpdesk server could reach.
Is this the first AI-agent breach?
No. The Cloud Security Alliance describes it as the second publicly documented autonomous AI-agent breach in under three months, after the Hugging Face intrusion in July 2026.

My Take

The symbolism is hard to miss. The organization that spends its days warning others about exposed systems was breached through its own helpdesk, by an attacker that probably was not human. Credit to DIVD for disclosing it quickly and in detail. That is exactly what we ask of everyone else, and it is not easy when you are the victim.

What worries me most is not that an AI agent hacked a security organization. It is that it did so badly and still won. It collided with its own attacks and wrote down its reasoning, and it still reached root in seconds. Today's agents are noisy. Tomorrow's will be quiet.

Most incident response plans still assume attacks unfold over hours or days, with time for a human to look at an alert. That assumption just expired. The teams that invest now in segmentation, behavior-based detection and automated containment will be ready. The others will learn from their own incident report.

So here is the question I will leave with you: is your incident response built for an attack that is over in seconds?

Sources:

This article is published for information purposes only, to help defenders understand a real AI-driven intrusion and protect their own systems. It contains no exploitation details.

Related Articles:

Kodjo Apedoh

About the Author

Kodjo Apedoh

Network Engineer & AI Entrepreneur

Founder of TechVernia & SankaraShield. Certified Network Security Engineer with 4+ years of experience specializing in network automation (Python), AI tools research, and advanced security implementations. Also builds iOS and Android applications. Holds certifications from Palo Alto Networks, Fortinet, and Cisco. Based in Arlington, Virginia.

Connect on LinkedIn →