The report
The Dutch security foundation DIVD says an attacker entered its infrastructure through two Zammad zero-days on September 21, 2026. Its rolling first-party disclosure added the identity of the ARTEX agentic pentest framework on October 31. The responder describes more than 450,000 requests, more than 2,500 scripts and a twelve-minute path from initial access to root. Those are victim-reported operational observations, not independent evidence that each request succeeded.
DIVD confirms partial exfiltration of volunteer and helpdesk data, while distinguishing other systems and categories that remain under investigation. The two flaws are tracked as CVE-2026-102489 and CVE-2026-102490, fixed in Zammad 7.2 and 7.2.2 respectively. The canonical Case keeps the September intrusion and all later disclosures together.
The ARTEX framework helps explain the attack's speed and repetitive structure, but the report also documents noisy mistakes and uncertain operator attribution. DIVD cautions that use of ARTEX elsewhere does not mean other incidents had the same actor. Compare Acronis BOSSAI: both studies support real offensive agent workflows while resisting unsupported claims of fully independent, uniformly successful AI hacking.
Two Zammad defects with different reachability
The vendor's Zammad advisory limits the remote exploitability of CVE-2026-102489 to older runtime conditions, while later packages received additional hardening. CVE-2026-102490 concerns a separate local privilege-escalation route affecting relevant packaged installations and addressed in 7.2.2. Those are not interchangeable fixes or proof that every live Zammad server exposed the same two-step attack. DIVD's own record connects the specific September compromise to ARTEX job logs, a non-expiring administrative API token and partial exfiltration. It does not show that the exact same entry point was used in every independently reported ARTEX campaign. The framework entity collects reusable software context without conflating operators.
Why the victim's own telemetry is stronger evidence
A claim that an attacker used AI can be based on branding, a script comment or a vendor's inference. DIVD's October 31 account instead identifies specific ARTEX job markers in systems it investigated, alongside a durable administrative API token and large volumes of commands. The organization also recorded failed and noisy activity. Those artifacts support the participation of an agentic framework in this single compromise; they do not prove every decision was delegated or that the actor behind unrelated ARTEX reports is the same. The durable ARTEX entity connects independent episodes by shared software, while the DIVD Case retains its separate victim, vulnerability and data-access evidence.