Field noteNov 16, 20266 min read1 source

ExtraHop counts 3,447 internet-exposed ARTEX hosts, more than double Censys’s 1,465

In one minute

ExtraHop’s Reveal Labs reports finding 3,447 ARTEX hosts and describes controlled tests of the framework’s agentic attack workflow. The count is higher than earlier published figures, but the measures are not directly comparable; the more useful operational point is how defenders can detect automated attack behavior without depending on ARTEX-specific signatures.

What changed: ExtraHop counts 3,447 internet-exposed ARTEX hosts, more than double Censys’s earlier count of 1,465.

What happened

ExtraHop published its analysis on November 15, 2026. It says the FBI issued a FLASH alert on November 13 warning that threat actors were using ARTEX, an open-source penetration-testing platform released on GitHub in July 2026. The framework uses large language models to translate objectives into commands and adjust its approach based on results. Its developer removed the original repository on October 8, but ExtraHop says forks and copies remained publicly available.

The post places ARTEX in a series of observed campaigns. Google Threat Intelligence Group saw threat actors use it as early as August 2026. CrowdStrike attributed a late-September campaign against 16 South Korean financial firms to a likely Chinese-speaking, financially motivated actor that used DeepSeek as its primary model. Microsoft reported Storm-0940 using ARTEX against hundreds of organizations and making automated authentication attempts against more than 2,250 Microsoft 365 tenants with compromised ConnectWise Cloud Backup credentials. Microsoft also reported Storm-3358 using ARTEX against financial institutions from October. ExtraHop says these examples show the framework can support multiple intrusion stages; the underlying techniques are largely established ones.

For its lab work, Reveal Labs used ScopeWeaver, an English-localization, rebranded ARTEX fork, and ScopeSentry, an optional reconnaissance tool. Operators define a target and objective; a planner breaks the objective into intents assigned to workers, which can run concurrently. Workers record findings in PostgreSQL, which inform the next planning and execution cycle. ScopeWeaver can use locally hosted models or cloud APIs, with multiple models configured as fallbacks. Operators can customize instructions, model choice and tools.

In simulations against Windows hosts, the team observed attempts involving SSH, RDP, Nmap, WinRM and PsExec. One task succeeded in changing a Windows registry value; another was reported as successful by ScopeWeaver after a connectivity check, despite no remote shell being established. ExtraHop says its RevealX network detection product generated behavioral detections around these activities, including scanning, remote-service attempts, plaintext authentication and use of dual-use tools. The tests were controlled simulations, not evidence that the framework successfully compromised an external target.

ExtraHop reports identifying 3,447 ARTEX hosts; 9.5% contained files associated with at least one other agentic attack framework, most often CyberStrike AI. The company excludes ARTEX derivatives and non-agentic tools from that analysis and says the result may therefore underestimate additional offensive tooling. It contrasts its count with Censys’s 1,465 publicly exposed consoles and Check Point Research’s 1,608 active instances, which it cites as separate findings.

The defensive guidance is to focus on behavior and correlated activity. ExtraHop observed suspicious user-agent strings and, when TLS session keys were available, JSON payloads with function names such as list_findings and shell_send. It also saw idle workers send keep-alives about every five seconds. The post cautions that these clues are not unique to ARTEX and that internal agent traffic may never cross a victim’s monitored network. It suggests correlating scanning and exploitation within short windows and reviewing unusual outbound connections after exploitation, alongside broader behavioral detection coverage.

What the archive adds

New Runtime’s earlier coverage reported Censys’s finding of 1,465 publicly exposed ARTEX consoles in a November 7 Field Note. ExtraHop’s 3,447-host estimate is a new fact on the same story, not a direct revision: the organizations describe different measurements. Earlier coverage also examined ARTEX-fork defenses and exposed attack panels, adding context about the ecosystem around the framework.

Trend signal

This supports agent security moving to runtime boundaries: ExtraHop’s tests emphasize observing network actions and technique sequences even when the specific framework or tool changes. The reported detections are useful signals, not proof of attribution; the company itself notes that tool-specific characteristics can be changed and that behavior-based coverage remains important.

Why now

The analysis follows the FBI’s November 13 FLASH and recent reporting by Microsoft and other researchers on ARTEX use in campaigns. ExtraHop’s lab tests connect that threat reporting to concrete network behaviors defenders can investigate.

Connections to other events

New Runtime’s earlier coverage also reported that an ARTEX fork used layered honeypot detection and anti-trap defenses, and that exposed Cairn, ARTEX and PentAGI consoles had reportedly been hijacked for credential theft and malware deployment. ExtraHop’s findings meet those stories at the intersection of framework design, exposed infrastructure and defender visibility; they do not establish that the 3,447 hosts were compromised.

Why it matters

For security teams, an agentic attack framework changes tempo more than it changes the basic vocabulary of an intrusion. ExtraHop cites observed attacks that advanced from targeting to privileged access in under three hours, while its lab simulations show agents trying alternative tools after failures. That can compress the time available for a human analyst to connect a scan, authentication attempt and subsequent action.

The practical implication is to build detections around sequences and outcomes rather than rely on a list of ARTEX identifiers. Short-window correlation can help surface automated trial-and-error activity, while broad coverage of scanning, exploitation, authentication and remote execution techniques can remain useful if attackers switch frameworks. Teams should treat agent access to telemetry as an investigation aid, not as a substitute for validating alerts and their evidence.

Limits and uncertainty

ExtraHop’s host count cannot be reconciled directly with Censys’s and Check Point’s figures from the descriptions in the announcement: their definitions, scan timing and methods are not shown as equivalent. The count is not a count of confirmed malicious deployments, compromised hosts or active attacks. ExtraHop says its analysis excludes ARTEX derivatives and non-agentic tools and may undercount additional tooling.

The lab used a rebranded fork and controlled targets; it demonstrates behaviors the researchers observed, not the success rate of real-world attacks. Some network indicators may be changed, shared by other tools or invisible to defenders when framework traffic stays outside their network. The post does not establish how often the suggested correlations produce true positives or false positives in production.

Skill to build now

Create one hunt that looks for scanning detections followed by exploitation on the same systems within five minutes, then inspect whether those systems make a first-time outbound connection soon afterward. ExtraHop gives these as example agent queries. Use the window as a starting point, adapt it to your telemetry and workload, and validate candidate sequences against underlying events before turning them into alerts.

Reader takeaway

ExtraHop’s larger ARTEX host count is a warning about uncertain measurement, not a definitive census of attackers. For defenders, the more durable lesson is to detect clustered behavior—rapid changes in technique, repeated access attempts and follow-on activity—because the framework can change while the underlying intrusion techniques remain recognizable.

Evidence / sources

  1. 1extrahop.com/blog/artext-inside-an-agentic-attack-framework
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicAgent SecurityExplore the agent-security topic hub.
  2. 02topicSecurityExplore the security topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract