Back to overview

    3,800 a day: Shadowserver now counts the servers running AI attack frameworks

    Since 15 November, Shadowserver has counted over 3,800 IP addresses per day running agentic AI attack frameworks such as ARTEX, CyberStrike and PentAGI, and reports hits with severity “high” to network operators. This is the measurement of what Google and Mandiant described: intrusions at machine speed, 27 actions in 23 seconds. What the frameworks do, why they are visible, how the attack shows up in your own network and which five layers remove the speed advantage.

    Published by NEOSEC
    KI-SicherheitAgentische AngriffeARTEXCyberStrikePentAGIShadowserverThreat IntelligenceEgressXIEMNIS2Mandiant
    Share

    As of 16 November 2026.

    On 15 November the Shadowserver Foundation announced it has added agentic AI attack frameworks to its daily device fingerprinting – specifically ARTEX, CyberStrike, PentestCode, VXControl and others. The result of the first scan: over 3,800 unique IP addresses per day running such infrastructure. The data feeds into the free Device Identification Report and is visible via the Shadowserver dashboard as a map and a time series; affected network operators receive hits flagged with severity “high”.

    This is the measurement of a development Google Threat Intelligence and Mandiant described two weeks earlier: attackers no longer run intrusions by hand but have an agent framework execute them. What used to be hours of operator work – finding credentials, testing permissions, moving laterally, extracting data – now happens in minutes. The weaknesses stay the same. The speed does not.

    What an agentic attack framework is.

    ARTEX, the framework behind most of the observed activity, is a Go-based multi-agent system with a web console. The operator sets a target and an objective; six specialised sub-agents break that into individual steps and execute them via a language model’s API – either a commercial frontier model or an open model such as DeepSeek. The original author deleted the repository in October 2026 because of widespread abuse; the framework lives on in dozens of forks, plus a commercial variant “ARTEX PRO”.

    The pattern is not new – attackers have always repurposed pentest tools like Cobalt Strike or Metasploit. What is new is that the orchestration itself is automated. Mandiant measured what that looks like in real incidents:

    • In one operation, 27 individual actions in 23 seconds – rapid probing of AWS permissions across multiple security groups. For a human operator this is impossible.
    • An actor tracked by Wiz (tracked by Microsoft as STORM-0940) moved from the first compromised account to deep system access in around four hours typically – in one case fifteen minutes.
    • The same speed creates noise: in one case a single attempt to write a JSP web shell via a SQL API left 240 log events, because several sub-agents overwrote one another.

    That last detail is the good news. The frameworks trade stealth for speed. They are fast but loud – and that is exactly why they are detectable at all.

    Why Shadowserver can see the infrastructure.

    The 3,800 IP addresses per day are not the victims but the attacker infrastructure: servers on which such a framework is installed and whose web console or API port faces the internet. The tools give themselves away through default configurations almost nobody changes:

    FrameworkFingerprint (example)
    ARTEXConsole on port 8787, path /api/auth/status reveals unconfigured state; cookie artex_token; self-updater calls with user agent artex-selfupdate, plus artex-egress-probe/1.0 and artex-enrich/1.0
    CyberStrikeHTTP title “CyberStrike”; user agent CyberStrikeAI/1.7.4 for scans, CyberStrikeAI-WebShell/1.0 when controlling planted web shells
    PentAGI (VXControl)Beacon to update.pentagi.com every three hours; service port open in many setups
    HexStrikeTools API hardcoded to 0.0.0.0:8888, unauthenticated

    For an attacker an exposed console is reckless – it lets any passer-by take over the running operation. For defenders it is a gift: Shadowserver scans it, attributes it to the operating organisation and reports it to the responsible network operator.

    Two readings of the report matter for you. First: if an IP from your address range appears in your Shadowserver report with such a fingerprint and you do not expect it, someone in your network is running an attack framework – a compromised server, a shadow project, or an in-house pentest IT does not know about. In all three cases, checking is mandatory. Second, and the more common case: the infrastructure is elsewhere, but its traffic lands on you. For that you need your own detection.

    How the attack shows up in your own network.

    Because the frameworks work with valid credentials and permitted functions (living off the land), no signature helps. The lever is behaviour and speed. Mandiant names the same places where any human attacker becomes visible too – only that the agent sweeps them at a per-second rate:

    PlaceWhat stands outImplementation in XIEM
    Identities / control planeBursts of management-API calls, permission tests with DryRun, rights enumeration in seconds; a service account logging in from a hosting network instead of its usual rangeCloud audit logs (Entra, AWS CloudTrail, GCP) into the SIEM; rule on API rate per identity and on new origin ASNs
    Data planeA service account that reads ten objects suddenly reads thousands; unusual write-read-delete sequences in storage bucketsAnalyse storage access logs, baseline per account, alert on volume jumps
    EgressA server beacons to changing webhook.site URLs (491 rotating in one case), to OAST services, tunnels or directly to LLM APIsNetFlow and proxy logs; rule “production server contacts new external destinations”, especially webhook.site and AI platforms
    HostFiles prefixed artex-probe-*, task folder /opt/artex/, incrementally named JSP web shells (cl1217.jsp, f1219a.jsp), accounts with ArtexProbe in the nameWazuh FIM on web and temp directories; Sysmon on curl calls that fetch PowerShell
    East-westA suddenly exposed web server initiating Kerberos, SMB, LDAP, WinRM into the internal networkHost and firewall logs – the same pattern as in the post on north-south and east-west

    The five layers that remove the speed advantage.

    Mandiant’s recommendation is soberingly unspectacular: the same fundamentals that have applied for years. The difference is that an agent framework finds the gaps between them in minutes where a human needed hours. Those who have the layers force the agent into exactly the noise that gives it away.

    1. Protect the control plane. Admin APIs, consoles, CI/CD, hypervisors, backup and deployment interfaces off the internet. The Azure Kudu/SCM interface was the entry point in several incidents.
    2. Restrict identities and keys. Replace long-lived secrets with short-lived, workload-bound credentials (managed identities, workload identity federation, OIDC instead of static keys). A key that works from anywhere turns a single find into full access instantly.
    3. Reduce attack surface and paths. Segment internet applications, database accounts without OS rights, no secrets in environment variables, no detailed error messages to the outside – a stack trace is fuel for the agent’s next attempt.
    4. Harden applications. Put rate and recipient limits on legitimate functions. In one incident ARTEX forced an application’s notification system to send test emails to over 1,300 internal users – a trusted function abused for mass distribution.
    5. Monitor behaviour and volume. Baseline per identity and workload, alert on jumps in API rate, data egress and new egress destinations. Detection is the layer that catches what the other four leave open.

    Context.

    GTIG places the current state at “stage 5” of a seven-stage scale: several phases of an intrusion already run without a human in the loop, but crudely, loudly and detectably. GTIG expects the next stage – broadly scattered, unsophisticated attacks with no operator at all – within one to three months. For the Mittelstand this does not mean attacks get cleverer. It means there are more of them and they run faster, because the staffing bottleneck on the attacker side falls away. An opportunistic actor who used to manage one target a week then manages twenty a day.

    That Shadowserver now counts and reports this infrastructure daily is the fitting answer on the defender side: visibility into the tools before they are pointed at your network. For entities subject to NIS2 the circle closes to Article 21: the Shadowserver report for your own networks and domains is free, and a process that investigates a “high” hit within hours is exactly the detection and response capability the law requires.

    What to do now.

    1. Subscribe to the Shadowserver report if you have not – free for your own IP ranges and domains. The Device Identification Report now contains the framework hits.
    2. Take “high” hits on your own infrastructure seriously. An unexpected ARTEX or CyberStrike signature from your network is an incident, not a false positive – isolate the server, check who runs the framework.
    3. Sharpen egress detection. Rules on webhook.site, OAST services and direct LLM-API calls from production servers. That is the most reliable single indicator of agentic activity.
    4. Review control plane and identities. Which admin interfaces face the internet, which credentials are long-lived and broadly permitted? Those are the paths an agent sweeps in minutes.
    5. Establish a baseline. No normal picture, no alert on deviation. Collect identity, control-plane, storage and egress telemetry centrally and set an expectation per account.

    Sources

    Affected products