Skip to content
AI Security Wire

Published

- 7 min read

By

FBI FLASH on ARTEX: The Operators, Their Playbook and Exposed Consoles

img of FBI FLASH on ARTEX: The Operators, Their Playbook and Exposed Consoles

The FBI has published a FLASH on ARTEX, the open-source LLM-driven penetration testing framework that attackers have been running against banks, governments and technology firms since at least August 2026. The FLASH (FLASH-20261113-001, TLP:CLEAR, 13 November 2026) does not name a single actor. Its subject is the tool and the crowd using it, which is why we treat it as an actor-ecosystem profile rather than a campaign report.

This article draws on the FBI FLASH, Amazon Threat Intelligence’s write-up (13 November 2026) of ARTEX records captured by its MadPot sensors, and, for specific figures, the Microsoft and Censys reports the FLASH cites. The FLASH names no victim organisations, only ConnectWise as the compromised vendor. Attribution below is the sources’ wording, not ours.

Who the FBI says is using it

The FLASH’s summary says “a variety of malicious cyber actors” have used ARTEX since at least August 2026 against financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal organisations. It names two campaigns:

  • Financially motivated attacks on South Korean financial institutions, exfiltrating personal information on approximately 67,000 individuals.
  • A supply chain compromise of ConnectWise Cloud Backup customers that, in the FLASH’s words, enabled access to more than 2,250 downstream tenants. Microsoft’s 21 October report, which the FLASH cites, words it differently: after the unauthorised access, Storm-0940 used ARTEX “to attempt to sign-in to more than 2,250 tenants”, and its infographic counts 327 accounts across eight organisations.

Microsoft calls Storm-0940 “a China-state aligned actor” active since 2021. The FLASH’s IP table also assigns indicators to “Storm-3358”, citing a separate Microsoft report (November 2026). We rely on the FLASH for those and say no more about Storm-3358.

Amazon Threat Intelligence gives a regional picture. It assesses “with moderate confidence” that ARTEX was used in the campaigns it observed, from prompt references to ARTEX tooling and other artifacts. It also assesses with moderate confidence that these are “disparate Chinese-nexus threat groups”, citing simplified Chinese in prompts, China time-zone configurations, targeting consistent with Chinese interests and tooling predominantly used by Chinese-nexus actors. AWS says sample instructions set red lines such as excluding mainland China targets and government domains, and required outbound traffic to go through proxies.

Our assessment: the sources are consistent with a tool shared by many operators, not one group’s capability. We found no state sponsor named in the FLASH body or the AWS post as of 17 November 2026, but the Microsoft report the FLASH cites for the ConnectWise campaign does name a China-state aligned actor, so this is not a purely criminal tool. Neither the FLASH nor AWS says the groups coordinate.

The tool and its supply

Per the FLASH, ARTEX is an autonomous, multi-agent penetration testing platform written in Go. It turns goals described in natural language into machine commands run by specialised agents. It ships without its own model and connects to external providers. The FLASH lists Anthropic Claude and OpenAI GPT alongside DeepSeek, GLM, Grok and Kimi, and says analysis of attacker-controlled instances shows Chinese models are most prevalent, with DeepSeek Flash family models dominating. It gives the reasons as capability, low cost and “lack of observable guardrails when framed for penetration testing”.

Supply points the FLASH makes:

  • The developer, “Autumn-27”, released ARTEX on GitHub in July 2026 and made the repository unavailable on 8 October 2026 after reports of abuse. It reached an estimated 1,500 to 2,000 GitHub stars within 10 weeks.
  • The framework persists through community forks, localised variants, archived copies and a commercial version, ARTEX PRO.
  • The FLASH says that as of early November 2026 researchers found over 1,600 publicly exposed ARTEX console instances. The Censys report it cites counts 1,465 hosts as of 4 November, and about 815 distinct deployments after excluding two identically configured hosting blocks; we use Censys’s numbers.

What operators do with it

AWS’s MadPot sensors posed as popular LLM inference APIs. ARTEX agents sent them prompts that carried their own campaign history, so AWS saw the operators’ instructions. Amazon Threat Intelligence did not observe the underlying activity as it happened and says it cannot verify the outcomes the agents reported. With that caveat, AWS reports:

  • Operators steer agents with standing instruction prompts that set objectives, targeting rules and constraints.
  • Most of the activity used ARTEX to divide work across planner, worker, verification and compression agents sharing a persistent knowledge graph.
  • One credential-theft operation worked through a fixed sequence of countries, prioritised enterprise production sites, and graded each credential: L1 a lead, L2 a confirmed working API call, L3 quantified impact such as readable data or spendable balance. Keys exposed publicly for more than 30 days were treated as low-value inventory.
  • Targets included LLM gateways and AI application builders (One API, New API, LiteLLM, Dify, Langflow, Open WebUI and others). Stolen keys were fed back into the framework’s model gateway to pay for further hunting.
  • A second campaign tried to pass cryptocurrency exchange KYC checks with real-time deepfakes. In the records AWS saw, one vendor’s server-side verdict rejected the deepfake in every recorded attempt.

The FLASH maps observed ARTEX activity to ATT&CK. Condensed:

PhaseObserved activity (FBI)
ReconSub-second capability checks on cloud credentials, DryRun permission tests, 500+ request directory fuzzing
Initial accessUnsigned SAML assertions, XXE, SSRF, anonymous upload, SQL injection, hardcoded admin passwords in JavaScript bundles, stolen VPN configurations, third-party vendor credentials via B2B federation
FootholdJSP and PHP web shells, Kudu API abuse, open-source RATs (WISP, vshell, AdaptixC2), new SSH keys and Elastic IPs, credentials added to OAuth apps
Privilege escalationKerberoasting, DCSync, resource-based constrained delegation, cloud metadata queries via XXE, forged admin tokens
Internal reconIterative PowerShell enumeration of over 110,000 AD objects, LAPS reads, inventory of endpoints lacking CrowdStrike Falcon
Lateral movementAttempted lateral movement to domain controllers using Pass-the-Ticket, WinRM, Kerberos relay, Pass-the-Hash (FLASH: “blocked by endpoint protection”)
Exfiltration prepMass blob reads, EC2 boot scripts, output to rotating webhook.site endpoints, new 100GB volumes
C2Residential proxies, Tor, Chisel, SoftEther; BLOODALCHEMY talked to C2 domains

The FLASH also says ARTEX is good at “peripheral systems receiving less security attention”: partner and supplier portals, mobile backends, sales-support systems and legacy services.

Indicators

The FBI says to vet indicators before blocking, because cloud addresses get reassigned. All values below are defanged and come from the FLASH’s tables, with its attributed sources.

  • 38.244.50[.]120: ARTEX host and open directory, South Korea financial campaign (CrowdStrike, Censys).
  • xcai[.]pro: LLM API reseller, same campaign (CrowdStrike).
  • 199.19.95[.]66, 199.19.95[.]67, 199.19.95[.]68: WISP C2 nodes, Storm-3358 campaign, active as of 6 November 2026 (Microsoft).
  • 91.92.240[.]238, 85.11.168[.]45, 178.16.54[.]129, 31.77.229[.]35: ARTEX servers, same campaign and date (Microsoft).
  • 85.11.167[.]52: C2 server, same campaign and date (Microsoft).
  • 144.172.64[.]60, 144.172.64[.]99, 144.172.64[.]128: ConnectWise campaign sign-ins and Graph API activity (Microsoft).

The FLASH also lists nine proxy IPs from the South Korean campaign and three transient VPN IPs; see the FLASH for those.

What defenders can do

The FBI’s recommendations focus on exposure and identity: inventory peripheral internet-facing apps, restrict Kudu/SCM and other management interfaces to trusted networks, apply least privilege to multi-tenant app permissions, rate-limit sensitive APIs and alert on sequential record lookups. For ConnectWise customers, it says to review every Microsoft 365 tenant where affected ConnectWise applications are or were authorised, searching by application ID rather than display name, and to keep affected connections disabled until ConnectWise gives formal restoration instructions.

Detection ideas: the FLASH’s Zscaler-sourced indicators include unexpected egress to api.anthropic.com, api.deepseek.com, open.bigmodel.cn or api.x.ai. Its reporting list also asks for evidence of self-hosted LLM gateway or reseller software (CLIProxyAPI, New API) in egress logs, and sub-second blob write/delete tests or ARM deployment dry-runs across resource types. Our caution: egress to those endpoints will fire on legitimate AI use in many environments, so scope it to servers and identities that should never call an LLM.

Our assessment: the notable shift is the attacker’s cost model. AWS reports, from the agents’ own records and without being able to verify outcomes, stolen AI keys paying for further key theft, and the FBI describes a framework that persists after takedown through forks and a commercial version. Takedown of the repository did not remove capability, which is the point defenders should plan around.

References

Frequently Asked Questions

Who is behind ARTEX attacks, according to the FBI?
The FBI does not name one actor. Its FLASH of 13 November 2026 says that since at least August 2026 'a variety of malicious cyber actors' have used the ARTEX framework against organisations in many sectors. Amazon Threat Intelligence separately assesses with moderate confidence that the operators it saw are disparate Chinese-nexus groups, based on simplified Chinese prompts, China time-zone settings and targeting consistent with Chinese interests.
Is ARTEX still available after its developer took it down?
Yes, according to the FBI. The developer made the GitHub repository unavailable on 8 October 2026 after reports of abuse, but the FLASH says the framework still circulates through community forks, localised variations, archived copies and a commercial version called ARTEX PRO. The FLASH says researchers found over 1,600 publicly exposed ARTEX consoles in early November 2026; the Censys report it cites counts 1,465 hosts as of 4 November, or about 815 distinct deployments after excluding two identically configured hosting blocks.
How fast can an ARTEX intrusion move?
The FBI says some intrusions progressed from initial targeting to privileged access in under three hours. The FLASH attributes the figure to Microsoft's October 21 write-up of the ConnectWise campaign, which says 'multiple instances' took less than three hours from initial targeting to privileged access.
What should defenders look for first?
The FLASH's network and behavioural indicators, sourced to Zscaler, are outbound calls to LLM API endpoints (api.anthropic.com, api.deepseek.com, open.bigmodel.cn, api.x.ai), high-frequency sequential enumeration of login and search endpoints, and .aspx or .jsp web shells in unexpected paths. For the API calls it says to investigate unexpected ones alongside other suspicious activity, and for all indicators it says to investigate or vet them before taking action such as blocking.