ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure

Authors: Cloud Security Alliance AI Safety Initiative
Published: 2026-11-16

Categories: Agentic AI Threat Intelligence
Download PDF
Download Markdown

Key Takeaways

ARTEX, an open-source multi-agent penetration testing framework released on GitHub in July 2026, is a widely documented example of an openly published red-team tool being adopted as general-purpose attack infrastructure. On October 8, 2026, after reports of abuse, its original developer announced that the project would become closed-source, and the GitHub repository and official Docker image were taken down shortly afterward. In ARTEX’s case, that withdrawal did little to limit availability, because the codebase had already been forked widely. Forks, translations, rebrands, and Docker images remain in circulation, and Wiz reported that the project had more than 12,000 forks by the time the original repository was withdrawn [3][4]. The codebase now exists in many versions, maintained by many parties with different intentions.

The threat activity is broad and well corroborated. Independent reporting from Google Threat Intelligence Group (GTIG), Wiz, CrowdStrike, Amazon, and others attributes ARTEX use since late August 2026 to multiple threat clusters, spanning PRC-nexus espionage actors and financially motivated groups [1][4][6][7]. Some of these clusters overlap with one another: GTIG’s UNC7228 and the actor Wiz tracked both overlap Microsoft’s Storm-0940. The infrastructure behind this activity is also visible. Censys counted 1,465 hosts exposing ARTEX consoles as of early November 2026, and 10.6% of them also hosted other threat tooling such as Vshell, Cobalt Strike, and Adaptix C2 [5].

The fork ecosystem is itself a supply-chain hazard. A trojanized ARTEX Docker image gave anyone who knew its trigger an unauthenticated root shell, which shows that operators of these tools can also become their victims [3][9]. Because ARTEX is AGPL-licensed, freely forkable, and model-agnostic, defenders should treat it as a commodity capability class rather than a single tool to signature. Detection should therefore focus on behavior, identity, and control-plane telemetry rather than on file hashes or repository names [1][2].

Background

ARTEX is a Go-based, multi-agent offensive framework with a Next.js web console and a PostgreSQL backing store. Public analyses of its README and forks describe a set of cooperating agent roles. A goals agent and a planner break an operator objective into discrete steps, worker agents carry out those steps concurrently with conventional security tools, and a main agent keeps a human-in-the-loop checkpoint [2][3]. The framework maintains a shared asset graph and a per-task exploration graph, records findings so they feed the next planning cycle, and supports MCP servers, reusable “skills,” a traffic-recording MITM proxy, and a one-click Docker installer that exposes the console on port 8787 by default [3][5]. Operators can attach local or cloud-hosted large language models with fallbacks, which means the framework does not depend on any single model provider [2]. In GTIG’s summary, ARTEX “abstracts the attack lifecycle” across its specialized sub-agents [1].

The project’s history moved quickly. According to Ostorlab’s reconstruction of public reporting, ARTEX was published by a developer reportedly based in China under the AGPL-3.0 license and won a red-team challenge in September 2026 [3]. Separately, beginning in late September, South Korean financial institutions reported intrusions that CrowdStrike later tied to an ARTEX console [3]. On October 8, the author posted a notice that the tool had been abused, that it would become closed-source, and that no further versions would be released. Within days, the GitHub repository and the official Docker image were gone [3]. An independent, unaffiliated field guide that tracks the ecosystem lists the last open upstream release as v0.3.15 and catalogs 49 notable public copies, divided into maintained successors, translations, mirrors, archives, derivatives, offensive builds, and rebuilds [8]. Its smaller count measures curated, distinguishable lineages, while Wiz’s figure of more than 12,000 measures raw GitHub fork events. The two numbers describe different things and are not in conflict.

Named derivatives illustrate how varied the ecosystem has become. GTIG identifies ScopeWeaver, an English-localized rebrand, and a commercial “ARTEX PRO” edition [1]. Censys observed consoles branded “ARTEX MX” and “ARTEX v2 · Agent Security Workbench” [5]. Ostorlab documents AntiFox, a fork of ARTEX 0.1.1 that reached version 2.0 and was used in AWS credential-theft intrusions [3]. Some maintained successors have advanced well beyond the last upstream release: according to the independent guide, active forks have reached v0.3.29, v0.4.5, and v0.5.1, with commits as recent as November 2026 [8]. At this point the codebase is effectively maintained by a decentralized community that the original author no longer controls.

Security Analysis

From Tool to Commodity Infrastructure

ARTEX deserves attention less because of any single technical capability than because of how widely it is available. Hexstrike, CAIRN, PentAGI, and CyberStrikeAI are comparable frameworks, and GTIG and ExtraHop both treat ARTEX as one member of a growing class rather than a unique threat [1][2]. ExtraHop cites GitGuardian’s count of roughly 45 other open-source agentic attack frameworks. ExtraHop’s own analysis found that 9.5% of 3,447 ARTEX hosts also carried files from at least one other agentic framework, most commonly CyberStrike AI [2]. These co-installations suggest that operators treat agentic frameworks as interchangeable components of a toolkit, a pattern that resembles the earlier spread of commodity C2 frameworks such as Cobalt Strike.

The Censys scan supports this interpretation. Of the 1,465 exposed ARTEX hosts, about 650 sat in two identically configured hosting blocks. Censys estimates roughly 815 distinct-looking deployments and attributes the uniformity to provisioning templates rather than hundreds of independent operators [5]. Censys found that 155 of these hosts (10.6%) also ran other threat tooling, including CyberStrikeAI on 70 hosts, PentAGI, HexStrike AI, Vshell, Cobalt Strike, Adaptix C2, FRP and NPS tunnels, and Gophish, and it identified 43 ARTEX hosts running LLM gateway or reseller software [5]. Censys cautions that some exposed consoles probably belong to authorized testers, and that detection based on page titles will miss renamed builds [5]. Even with those caveats, the data indicates an operational footprint of roughly 815 distinct deployments, co-located with conventional post-exploitation infrastructure.

The host counts from Censys and ExtraHop are not directly comparable. ExtraHop counted exposed directories operating as ARTEX hosts, while Censys counted consoles matched by page title and Next.js asset references as of November 3–4, 2026. Ostorlab summarizes the range as roughly 800 to 1,650 exposed consoles, depending on what each scanner counted [3]. The two co-installation rates (9.5% and 10.6%) likewise rest on different denominators and should be read as independent estimates rather than as a single trend.

Observed Campaigns

Public reporting on ARTEX abuse now spans several vendors and government sources, each with its own visibility and attribution method. GTIG reports that a diverse range of threat actors has adopted ARTEX since late August 2026 [1], and later reports have added actors, sectors, and geographies. The table below summarizes the principal public reports of ARTEX use by threat actors. Attribution and scope figures come from each reporting organization and have not all been independently corroborated.

Reporting source Actor / cluster Targeting and observed behavior
GTIG / Mandiant [1] UNC7228 (overlaps Microsoft’s Storm-0940); financially motivated extortion groups; other suspected PRC-nexus clusters UNC7228 targeted enterprise software, semiconductor, telecom, and industrial automation; the clusters collectively affected government and manufacturing as well. From first account to deep access in about four hours on average, and in one case 15 minutes
Wiz Research [4] Actor overlapping Storm-0940 IT and software vendors; agentic probing of Okta, Entra ID, and AWS Cognito; password spraying, account-recovery abuse, unsigned JWTs sent to more than 60 SSO route variants
CrowdStrike, via BleepingComputer [7] VAPOR PANDA More than 200 organizations in 14 countries; ARTEX used primarily with DeepSeek alongside a second commercial model
Amazon Threat Intelligence [6] Disparate Chinese-nexus groups (moderate confidence) Theft of AI-service credentials financed by stolen inference; deepfake attempts against cryptocurrency exchange KYC checks
ExtraHop summary of multiple reports [2] Storm-3358 (Cambodia-based) and others Financial institutions since October 2026, with RATs including Vshell and AdaptixC2

Two features stand out in the most detailed reports. GTIG emphasizes speed: it measured execution bursts of roughly 2 to 25 seconds, including one sequence of 27 events in 23 seconds, and recorded a single exploitation sequence that generated 240 log events [1]. Wiz emphasizes persistence through iteration, observing that most individual attempts failed, but the agent kept adapting until it found a weakness, and that sessions lasted 24 to 30 hours with new source IPs appearing within minutes of each block [4]. Together these findings suggest that controls which rely on an attacker giving up, or on an analyst triaging alerts before the next step, are less effective against agentic operators.

Government reporting reinforces the vendor findings. On November 13, 2026, the FBI issued FLASH-20261113-001, warning that threat actors use ARTEX to automate intrusions “at machine speed, from reconnaissance through data exfiltration” [14]. The alert, released as TLP:CLEAR, also dates threat-actor use of the framework to at least August 2026, consistent with GTIG’s timeline [1][14].

The Fork Ecosystem as a Supply-Chain Risk

Forking has introduced a risk that falls on the operators of ARTEX as well as its targets. Ostorlab reports that the tomicdd/artex Docker image, the most-starred copycat, added an unauthenticated login handler that executed shell commands as root. Docker removed the image on November 5, but 637 GitHub repositories still referenced it that day [3]. Ostorlab, citing BadByte, also reports roughly 34 Docker Hub namespaces publishing images named “artex,” with about 60,000 combined pulls across all namespaces; the reporting does not say how many of those pulls involved malicious images [3]. CSA’s earlier research note on this incident advises that any deployment built from the trojanized images should be treated as fully compromised and rebuilt [9].

This matters to enterprises, not only to attackers. Legitimate red teams, managed security providers, and internal security engineering groups are evaluating agentic pentest tools. With the upstream repository gone, someone searching for “ARTEX” is likely to land on whichever fork or image ranks highest, and the provenance of that artifact is unclear. A compromised offensive framework is a high-value foothold because it is deployed with broad network reach, privileged credentials for testing scope, and outbound connectivity to model APIs, and because security teams commonly exclude authorized testing activity from alerting. Organizations that run any agentic offensive tool should apply the same software-supply-chain diligence they use for production dependencies, and arguably more.

Agent Memory as an Intelligence Source and a Liability

Amazon’s MadPot findings point to a second-order exposure. From October 3 to November 5, 2026, MadPot sensors that imitated LLM chat-completions APIs received requests from ARTEX agents that had been pointed at them as inference backends [6]. Because agent frameworks resend conversation history, the resulting requests carried task context, system prompts, tool histories, knowledge-graph state, and reported findings. In effect, the captured traffic contained the operational record of the campaign [6]. For defenders, this means that inference traffic leaving an agentic attack framework is rich in intelligence and, where defenders or providers control or observe the inference endpoint, can be captured. For any organization running its own agents, the same mechanism shows that a misconfigured or hijacked model endpoint can leak an agent’s working memory. The lesson applies in both directions, and it extends to legitimate red-team and IT-automation agents.

The same campaign data reinforces the connection between agentic frameworks and LLMjacking. Amazon describes a self-sustaining operation in which stolen AI-service keys paid for the inference that drove further key hunting. Its targets included LLM gateways and AI application builders such as LiteLLM, Dify, Langflow, RAGFlow, Open WebUI, and Ollama [6]. AI infrastructure is therefore both a target of these frameworks and a resource they consume.

Detection Considerations

ARTEX produces some artifacts that can be fingerprinted. ExtraHop’s lab simulation of ScopeWeaver observed user-agent strings such as scopeweaver and artex-enrich/1.0, LLM API payloads containing function names such as list_findings, record_fact, and shell_send when TLS can be decrypted, and keep-alive messages from idle workers about every five seconds [2]. Censys fingerprints consoles by page title and Next.js asset references [5]. These indicators are fragile, however, because forks can rename strings, change ports, and rebuild front ends with little effort. Wiz cautions that most of the user agents it observed, including curl, python-requests, and Boto3, are generic and noisy when hunted alone [4].

Durable detection therefore depends on behavioral and volumetric signals rather than tool-specific strings. Useful signals include bursts of control-plane activity, new authentication sources, bulk data access, and connections to webhook, OAST, or tunneling services. The last of these is well illustrated by GTIG’s observation of one operator sending command output to 491 rotating webhook.site URLs [1].

Recommendations

Immediate Actions

Organizations should first eliminate the exposures agentic operators find fastest. Amazon and GTIG both stress the basics: remove default and hardcoded credentials, rotate secrets found in code and public artifacts, close unauthenticated administrative interfaces and inference endpoints, and patch internet-facing software, on the assumption that exposed weaknesses will be found quickly [1][6]. Security operations teams should load the published indicators from Wiz, Amazon, and the vendor detection packs referenced by GTIG, and should hunt identity provider logs for the password-spray and account-recovery abuse patterns Wiz describes [1][4][6]. Any team that has deployed ARTEX or a derivative for authorized testing should identify the exact fork and image used, check it against the trojanized images publicly reported, and rebuild from a verified source if provenance cannot be established [3][9].

Short-Term Mitigations

Over the next several weeks, organizations should build detections and controls that do not depend on tool names. GTIG’s five-layer guidance offers a practical structure for this work, summarized in the table below [1].

GTIG layer Recommended controls
1. Protect control planes Restrict administrative APIs, consoles, and CI/CD control planes to trusted networks, and require phishing-resistant MFA such as FIDO2 or passkeys.
2. Restrict identities and keys Replace long-lived secrets with short-lived, workload-bound credentials, and constrain where each identity can operate.
3. Harden configurations and reduce attack paths Run continuous vulnerability scanning across applications and infrastructure, and remove the weaknesses that give attackers a foothold.
4. Application protections Add rate limits and recipient caps to high-impact business APIs such as email, notification, and user-creation endpoints.
5. Detect behavioral and volumetric change Baseline identity, origin, and API volume, and alert on control-plane bursts and on connections to webhook, OAST, or tunneling services.

Beyond GTIG’s framework, AI gateways, model proxies, and inference endpoints should be brought under the same controls as other production systems, with spend and usage monitoring that can surface LLMjacking [6]. Teams responsible for SOC automation should also review response playbooks, because Wiz’s observation of new source IPs appearing within minutes of each block suggests that manual, IP-based containment will lag agentic adversaries [4].

Strategic Considerations

Over the longer term, the ARTEX case suggests that withdrawing an open-source offensive tool after it is abused has limited containment value once the tool has been forked widely. Security programs should plan on the assumption that capable agentic attack frameworks will stay freely available and will keep diverging into many variants. That argues for continuous, adversary-informed testing at machine speed, including authorized use of agentic red-team tools under enforced scope boundaries, verified provenance, and isolated execution environments. Organizations that develop or publish offensive AI tooling should consider release practices that address dual-use risk, such as gated distribution, signed releases, and secure-by-default consoles that do not listen on public interfaces. Procurement and third-party risk processes should also ask service providers which agentic offensive tools they run, where those tools come from, and how their memory and model traffic are protected.

CSA Resource Alignment

CSA’s whitepaper Autonomous Agentic AI Adversaries argued in June 2026 that autonomous agentic adversaries had moved from theory to operational reality [10]. ARTEX is a well-documented public example of the shift that whitepaper described. The whitepaper’s guidance on detecting machine-speed operations and strengthening identity and control planes is relevant to the campaign behavior described above. This note adds a factor the earlier paper did not address: the ARTEX case suggests that, once released and widely forked, an agentic attack capability can spread faster than a single maintainer or platform can withdraw it.

CSA’s research note LLMjacking Evolves: Stolen AI Compute as Attack Infrastructure described stolen AI credentials as a foundational resource for autonomous offensive frameworks [11]. Amazon’s MadPot observation of ARTEX agents financing their own inference through harvested keys is an example of the dynamic that note described. Organizations should read the two notes together when prioritizing protection of AI gateways and model credentials.

Two recent CSA research notes cover narrower parts of the ARTEX picture. Trojanized Offensive AI Tooling: The ARTEX Docker Backdoor analyzes the backdoored image in detail and provides remediation guidance for affected deployments [9]. Multi-Agent Offensive Frameworks in the Wild: CyberStrikeAI and ARTEX examines the Japan and Korea campaigns and the overlap between frameworks that Censys and ExtraHop later quantified [12]. This note places both in the wider context of the fork ecosystem.

For control implementation, the AI Controls Matrix (AICM) v1.1 provides the relevant baseline [13]. Its Identity and Access Management domain addresses the short-lived credential and phishing-resistant authentication measures recommended here. The Threat and Vulnerability Management and Logging and Monitoring domains support behavioral detection and rapid remediation. Its supply-chain and model-security controls apply both to protecting inference endpoints and to verifying the provenance of any agentic security tooling an organization deploys.

References

[1] Google Threat Intelligence Group. “ARTEX and Friends: Enter Agentic Offensive Frameworks.” Google Cloud Blog, October 30, 2026.

[2] Pyle, Taylor (ExtraHop Reveal Labs). “ARTEX: Inside an Agentic Attack Framework.” ExtraHop, November 15, 2026.

[3] Ostorlab Team. “ARTEX: How a Free AI Pentest Framework Became an Attacker Staple.” Ostorlab Blog, November 13, 2026.

[4] Federovsky, Sapir, and Rami McCarthy. “ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software Vendors.” Wiz Research, October 22, 2026.

[5] Cutler, Silas. “ARTEX: Agentic Asbestos.” Censys ARC, November 5, 2026.

[6] Moses, CJ. “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents.” AWS Security Blog, November 13, 2026.

[7] Toulas, Bill. “Chinese hackers use ARTEX AI in automated attacks against 200 orgs.” BleepingComputer, November 2, 2026.

[8] ARTEX.sh. “ARTEX.sh: The independent ARTEX field guide.” Unaffiliated community resource, accessed November 16, 2026.

[9] Cloud Security Alliance AI Safety Initiative. “Trojanized Offensive AI Tooling: The ARTEX Docker Backdoor.” CSA Labs, November 6, 2026.

[10] Cloud Security Alliance AI Safety Initiative. “Autonomous Agentic AI Adversaries.” CSA Labs, June 24, 2026.

[11] Cloud Security Alliance AI Safety Initiative. “LLMjacking Evolves: Stolen AI Compute as Attack Infrastructure.” CSA Labs, June 18, 2026.

[12] Cloud Security Alliance AI Safety Initiative. “Multi-Agent Offensive Frameworks in the Wild: CyberStrikeAI and ARTEX.” CSA Labs, October 28, 2026.

[13] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” Cloud Security Alliance, 2026.

[14] Federal Bureau of Investigation. “Threat Overview of ARTEX Autonomous Penetration Testing Framework.” FBI FLASH-20261113-001 (TLP:CLEAR), November 13, 2026.

← Back to Research Index