Insights

    AI Security

    ARTEX turns exposed web apps into an identity response problem

    Devence Lab

    · 2 min read

    Share
    ARTEX turns exposed web apps into an identity response problem
    Photograph · panumas nikhomkhai / Pexels

    Microsoft observed Storm-3358 using ARTEX against financial firms. The control point is the jump from public web flaws to cloud identity and persistence, not the tool's agent label.

    In October 2026, Microsoft Threat Intelligence observed Storm-3358 targeting financial organisations with ARTEX, an agentic offensive framework. Its November 6 report describes public web scanning, exploitation of anonymous file upload and server-side request forgery, and installation of remote access tools. The operational lesson lies in the transitions between those steps.

    An exposed application is only the first foothold. Microsoft also describes web shells, compromised credentials, abused cloud identities and data exfiltration. A defender who treats the initial web alert as a self-contained incident can miss the later identity and endpoint activity.

    Follow the attack across control planes

    Microsoft says ARTEX agents have used netcat and curl to download WISP, vshell and AdaptixC2 remote access tools. The same campaign used ordinary automation alongside ARTEX. That combination matters: the incident cannot be classified by whether a single request looks AI-generated. Investigators need a timeline linking web requests, process creation, outbound connections, new files and sign-ins.

    Start with internet-facing applications that permit file upload or server-side fetches. Correlate suspicious requests with shell execution by the web service account, subsequent curl or wget activity, and new outbound destinations. Then inspect credentials reachable from that host and the cloud identities that used them. Microsoft lists relevant endpoint and identity detections, but notes that some alerts can also fire on unrelated activity.

    The response boundary is the first credential the web foothold can use, not the first request the agent sends.

    Use behaviour before brittle signatures

    Microsoft reports that agentic scanning sends varied requests in a short period and quickly retries techniques. Default tool identifiers from Nmap, Python clients or headless Chromium can help triage, but they are easy to change. Its published IP indicators are useful for immediate hunting, not a durable boundary: the report itself warns that attackers rotate proxies and VPNs.

    The stronger test is whether the organisation can see a fast sequence from probing to exploitation to persistence across logs. Run a short hunt against recent web, endpoint and identity events; preserve the sequence, not just matching IPs. Review whether service accounts and cloud roles can be limited before the next foothold.

    Make deception measurable

    Microsoft says its tests with honeypots, decoys, tarpits and false success signals reduced traffic reaching real assets and generated high-confidence evidence. That is a useful design direction, but it is not a substitute for access control. Place decoys where an intruder would enumerate next, confirm that touching them produces an alert, and measure whether the alert reaches an analyst before any sensitive credential is used.

    For financial teams, the immediate change is practical: rehearse one end-to-end case that begins at a public app and ends at a cloud identity. If the two halves cannot be joined quickly, an agent can traverse the gap faster than the response process.

    Sources

    1. ARTEX abuse expands as cybercriminals target financial sector · Microsoft Threat Intelligence

    Written by the Devence Lab research team.

    Share