CERT-EU, the cybersecurity service for the EU institutions, has warned of a growing number of incidents in which attackers hand large parts of multistage intrusions to agentic AI frameworks. As reported by heise, the agency expects more frequent, faster and more scalable automated attacks with a higher success rate.
What happened
At the centre of the warning is Artex, an agent framework originally released for authorised penetration testing. Despite its “authorised use only” terms, it has become popular with criminals. Its author removed the repository in October 2026 after abuse was reported, but independent forks continue to circulate, and comparable tools such as Cairn exist. Microsoft had reported Artex-driven campaigns against financial organisations a few days earlier, and the abuse of Artex was also cited in the Chainstack intrusion disclosed in late October.
How it works
An operator supplies a high-level objective, for example obtaining domain admin rights. An orchestrator agent spawns sub-agents, assigns them discrete tasks, collects their results, plans the next step and tries alternatives when a technique fails. The agents drive well-known tooling, such as BloodHound to map Active Directory attack paths and Rubeus for Kerberoasting.
heise stresses that the techniques themselves are not new; they already underpin thousands of ransomware intrusions. Security researcher Florian Roth, who has predicted a wave of such attacks, argues that the change lies in speed and scale: less skilled operators can run many intrusions in parallel, altering the economics of attack.
Impact
Defenders should expect compressed timelines between initial access and domain compromise, more simultaneous campaigns and fewer manual mistakes that previously gave analysts time to respond. Weak identity hygiene in Active Directory becomes even more costly.
Recommendations
- Eliminate Kerberoastable service accounts with weak passwords; move to gMSAs and AES-only Kerberos.
- Run BloodHound against your own environment and remove the shortest paths to Tier 0.
- Alert on bursts of LDAP enumeration, ticket requests for many SPNs and rapid tool execution from a single host.
- Automate containment (isolating hosts, disabling accounts) so response keeps pace with machine-speed attacks.
- Expose MCP servers and monitoring stacks such as Grafana only with authentication and without default credentials.