What happened
On November 13, 2026, the FBI published FLASH-20261113-001 describing ARTEX-linked tactics, indicators and mitigations. ARTEX is an open-source multi-agent penetration-testing framework written in Go. The FBI lists Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi, but does not identify model versions used in the cited campaigns. Operators define high-level tactical goals, and ARTEX translates them into commands for specialised agents. Observed ARTEX activity automated reconnaissance, permission checks, exploitation, credential changes, persistence and exfiltration preparation.
The FBI says malicious use has occurred since at least August 2026, while Microsoft observed Storm-0940 using ARTEX in the ConnectWise campaign during October 2026. The FBI says some intrusions reached privileged access in under three hours and the ConnectWise compromise enabled access attempts against more than 2,250 downstream tenants. Microsoft says activity between October 10 and 14, 2026 produced more than 151,000 failed logons and peaked at more than 2,900 failures per minute.
Microsoft’s hunt identifies application IDs 506e08f9-6e9a-4da4-942b-a099a03644ae, 24205314-2b89-4e07-a2a4-579e898a5640 and 0de14c77-b9b0-4137-9360-34878a929827. Microsoft associates 144.172.64[.]60, 144.172.64[.]99 and 144.172.64[.]128 with ConnectWise application sign-ins and Graph API activity. Defenders can correlate EntraIdSpnSignInEvents with GraphAPIAuditEvents and remove the IP filter for broader application-ID hunting. Attribution posture: Microsoft attributes the ConnectWise campaign to the China-state-aligned actor Storm-0940, while the FBI says multiple actors have weaponised ARTEX.
Why this matters now
The leadership change is response latency. Existing playbooks often assume time to validate an alert, assemble cloud, identity and application owners, and then contain access. The FBI’s account shows some ARTEX-enabled intrusions reaching privileged access in under three hours, so high-confidence service-principal and application-control-plane alerts require pre-authorised containment paths.
Compromised application identities can authenticate without an interactive user sign-in. That weakens monitoring strategies centred on user MFA prompts or endpoint activity and makes application credentials, consent grants, Graph operations and privileged role modifications first-class incident signals.
ARTEX also broadens the economically viable target set. The FBI highlights partner portals, employee applications, legacy services, mobile backends and third-party integrations: systems that often have weaker ownership, testing and telemetry than flagship applications but can still expose trusted routes into cloud estates.
The decision for security leaders
Require cloud identity and SOC teams to run the application-ID hunt across current and formerly authorised ConnectWise integrations. Successful authentication must be correlated with Graph operations, credential changes and directory modifications before disposition.
Reduce the response threshold for high-privilege service-principal anomalies. Pre-authorise containment for unauthorised application credentials, owner additions, consent grants and privileged role changes so responders do not wait for a user-centric signal.
Expand attack-surface governance beyond flagship applications. Assign owners to partner portals, employee tools, mobile backends, legacy services and publishing interfaces, then restrict administrative access and add rate, volume and sequential-access detections. Require a documented response owner, application-identity scope, access evidence, containment decision and downstream cloud-impact assessment before closing the investigation.
Evidence of closure
- Tenant-by-tenant hunt records cover all three published application IDs.
- No unauthorised application credentials or roles remain, or each match has an incident disposition.
- External-application inventory assigns an owner and exposure state to every identified portal.
- Detection tests prove service-principal and Graph anomalies reach the on-call responder.
The Security.io assessment
The material development is not proof that an AI system independently selected targets. Operators supplied objectives and compromised access; ARTEX mechanically accelerated command generation and execution across multiple stages. Governance should therefore focus on identity, permissions, exposure and response speed rather than speculative claims about intelligence.
Microsoft cautions that an affected application or successful authentication does not by itself prove data access or exfiltration. Findings require correlation with Graph activity, directory changes, resource access and expected vendor infrastructure before declaring compromise.
The FBI’s infrastructure indicators are historically observed and may be reassigned or transient. Application IDs, behavioural patterns and identity changes provide stronger durable hunting pivots than blocking IP addresses without corroborating telemetry.
Questions for the morning meeting
- Can the SOC investigate service-principal activity within an attack window measured in hours?
- Which multi-tenant applications retain privileges that exceed their current business purpose?
- Are peripheral portals and publishing interfaces inventoried and owned?
- Can cloud teams preserve Graph and Entra evidence before removing malicious credentials?