On September 21, 2026, an attacker breached the Dutch Institute for Vulnerability Disclosure (DIVD) through their Zammad helpdesk. It took the attacker 12 minutes to get root access, including about a minute for the final privilege escalation. DIVD detected the intrusion seven minutes after the attacker got root. During their investigation, DIVD identified the attacker’s tooling as ARTEX, an open-source penetration-testing framework (DIVD, 2026).
I reviewed the source code of ARTEX v0.3.15 to understand how it differs from a general-purpose agent harness (Go module proxy, 2026). ARTEX provides task planning, shared findings and HTTP capture out of the box, so the operator doesn’t have to wire them together.
What attackers could already automate
Automation in red teaming predates modern LLMs. By 2014, an operator could run SpiderFoot to find documents associated with a target domain (Micallef, 2014).
Months before the DIVD breach, threat actors were already using AI agents. Unit 42 recovered a Hermes Agent session in which DeepSeek switched from Langflow to n8n after the first attempt failed, with no further operator input in the session record (Unit 42, 2026b). In July, Dream also observed multiple Hermes and OpenClaw agents in an attacker’s campaign (Dream Research Labs, 2026).
Inside ARTEX
In ARTEX, an orchestration agent handles the operator’s requests, and a planner can queue additional work as new discoveries are made and recorded. Multiple workers can claim work items from the same queue (ARTEX, 2026g; ARTEX, 2026h).
Suppose a worker finds a customer record on a helpdesk page. It can save what it saw and attach the request and response. A worker on a linked task can then read the finding without requesting the page again (ARTEX, 2026a; ARTEX, 2026i). If I saw that one request in the helpdesk log, I’d check which customer fields the page returned to that account.
ARTEX can retain credentials in captured HTTP headers and bodies (ARTEX, 2026d; ARTEX, 2026c). If the response exposed a token, I’d check the logs of the service that accepts it for subsequent use.
Time to privilege escalation
In November 2025, before ARTEX was released, Sysdig observed a suspected AI-assisted intrusion in which an attacker went from stealing exposed AWS credentials to running code with administrative privileges in eight minutes (Sysdig, 2026).
Where defenders can interrupt the intrusion
Monitor for privilege escalation, for example when an application or service account grants itself administrative privileges. Huntress saw the compromised Teams application assign itself Global Administrator. During the investigation, look for other role grants or credentials added by the attacker and revoke them (Huntress, 2026; Microsoft, 2026a).
Defenders can also automate parts of their response. For example, a playbook can disable an account that grants itself an unauthorized admin role while an analyst reviews the incident.
Defenders should also consider deletion protection for critical assets. In the Storm-3168 intrusion, the attacker deleted most of the targeted Azure storage accounts in about seven minutes. Resource locks or deletion protection blocked deletion on a few accounts (Microsoft, 2026b).
References
- ARTEX. v0.3.15 operator-facing agent and work requests. 2026.
- ARTEX. v0.3.15 planning and worker dispatch. 2026.
- ARTEX. v0.3.15 finding creation transaction. 2026.
- ARTEX. v0.3.15 related-task lookup. 2026.
- ARTEX. v0.3.15 HTTP snapshot and attachment schema. 2026.
- ARTEX. v0.3.15 HTTP body storage. 2026.
- DIVD. We got hacked, here’s everything we know. 2026.
- Dream Research Labs. Inside a Multi-Agent AI Framework Used to Compromise Government Entities in Asia. 2026.
- Go module proxy. Original ARTEX v0.3.15 release metadata. 2026.
- Huntress. AI, Scripts, Operator: Inside the ConnectWise Attack. 2026.
- Steve Micallef. SpiderFoot file discovery module, June 2014 revision. 2014.
- Microsoft. AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise. 2026.
- Microsoft. Storm-3168: Agentic-driven cloud attacks using compromised service principals. 2026.
- Sysdig. AI-assisted cloud intrusion achieves admin access in 8 minutes. 2026.
- Unit 42. Chinese-Speaking Threat Actor Harnesses AI Models for Autonomous Cyberattacks. 2026.