Skip to content
All articles
November 14, 2026 13 min read

FBI FLASH on ARTEX: How Attackers Turned an AI Pentesting Framework Into an Intrusion Engine, and How to Defend

Chris Rees

Chris Rees

25+ years in IT · Pluralsight author, 4.6/5 across 2,000+ ratings

FBI FLASH on ARTEX: How Attackers Turned an AI Pentesting Framework Into an Intrusion Engine, and How to Defend
Skillthropicwww.skillthropic.com

On November 13, 2026 the FBI published FLASH-20261113-001, a TLP:CLEAR warning about ARTEX, an open-source penetration testing framework that hands an intrusion to a team of AI agents. Give it a goal in plain English, connect it to a large language model, and it plans and runs the reconnaissance, exploitation, privilege escalation and data collection on its own. The FBI says threat actors have used it since at least August 2026 against financial services, government, defense, healthcare, energy, telecom and more, with some intrusions going from first targeting to privileged access in under three hours. This post breaks down how ARTEX works, what attackers actually did with it (mapped to MITRE ATT&CK), the indicators to hunt for, and what defenders should change, ending with three things you can do today.

Rather watch or listen? Here's the video version, about four and a half minutes. There's also a 13-minute podcast deep dive that goes further into the indicators, the hunting queries and the FBI's mitigations.

What is ARTEX?

ARTEX is an autonomous, multi-agent penetration testing platform written in Go. A developer using the handle "Autumn-27" released it on GitHub in July 2026 for authorized testing, and it collected an estimated 1,500 to 2,000 stars in ten weeks. After reports of abuse, the developer pulled the repository on October 8, 2026. It did not matter. The FBI notes that ARTEX keeps spreading through community forks, localized variants, archived copies and a commercial edition called ARTEX PRO, and that researchers counted more than 1,600 publicly exposed ARTEX consoles in early November, with the real number of deployments likely far higher.

The design is what makes it dangerous. ARTEX has no model of its own. It is an orchestrator: it takes a high-level objective, breaks it into tasks, and assigns them to specialized agents (the FLASH calls them "SI agents") that turn each task into real commands. The reasoning comes from whichever LLM the operator plugs in. The FBI lists frontier models (Anthropic Claude, OpenAI GPT) and open or Chinese-origin models (DeepSeek, GLM, Grok, Kimi). In attacker-run instances, Chinese models dominate, led by the DeepSeek Flash family, because they are cheap, capable and showed no observable guardrails when the request was framed as a penetration test. Operators often reach them through LLM API resellers and self-hosted gateways rather than a provider account in their own name.

How ARTEX works: an operator types an objective into the ARTEX console, a Go orchestrator that sends prompts to an external LLM API such as DeepSeek, GLM, Kimi, Grok, Claude or GPT, often through a reseller. The console runs four kinds of sub-agents in parallel, for recon, exploitation, cloud and Active Directory, against the target's attack surface. A band at the bottom shows where defenders can see it: egress to LLM APIs, the shape of the traffic, and cloud and identity telemetry. Operator types a goal in plain language, then waits objective ARTEX console Go orchestrator: plans, assigns tasks, retries prompts External LLM API DeepSeek (most used) GLM, Kimi, Grok, Claude, GPT often via a reseller Recon agent scans, fuzzes, maps perms Exploit agent SSRF, XXE, SQLi, SAML Cloud agent Kudu, IAM roles, blobs AD agent Kerberoast, DCSync YOUR ATTACK SURFACE partner and supplier portals · mobile backends · Kudu/SCM publishing endpoints · SAML sign-in · leaked cloud keys and VPN configs · legacy services · third-party integrations WHERE DEFENDERS CAN SEE IT Egress servers calling LLM APIs or unfamiliar LLM resellers Traffic shape bursts, sequential lookups, duplicate requests and retries Cloud and identity sub-second blob tests, ARM dry-runs, new app secrets
ARTEX is an orchestrator, not a model. The reasoning is rented from an LLM API, and the work is split across sub-agents that run in parallel. Each layer leaves something a defender can log.

How attackers are using ARTEX

The FLASH describes ARTEX less as a new exploit and more as a force multiplier. Nothing in its toolbox is novel: SQL injection, SSRF, Kerberoasting and web shells have all been documented for more than a decade. What changes is who can chain them and how fast. The FBI calls out five risks:

  • Accelerated multi-target operations. The agents automate the slow parts of an intrusion, enumeration and exploitation, so a small group can run several campaigns at once.
  • Overlooked attack surfaces. ARTEX is good at finding the systems nobody watches: partner and supplier portals, mobile backends, employee apps, sales-support tools, legacy services, third-party integrations, and apps a business unit stood up on its own.
  • Supply chain compromise. Agents map trust relationships between organizations and go through the weakest one. In one case they modified iOS app distribution infrastructure using exposed cloud credentials.
  • A lower barrier to entry. Operators with minimal skill can run intrusions that used to need an experienced team.
  • A compressed detection window. ARTEX is noisy and can trip alerts. The problem is that it can reach its objective before a defender has finished triaging them.

Two campaigns show the scale. A financially motivated actor used ARTEX against South Korean financial institutions and exfiltrated personal information on about 67,000 people. And a supply chain intrusion against ConnectWise Cloud Backup customers opened access to more than 2,250 downstream tenants, which is why a large share of the FLASH's mitigations are about Microsoft Entra and app consent.

How one vendor compromise fans out: ARTEX compromises a vendor's credentials or adds a credential to the vendor's multi-tenant OAuth application. That application already has admin consent in every customer tenant, so the attacker reaches Graph API data in each one, more than 2,250 tenants in the ConnectWise campaign. A defender check band says to search by application ID, review service principal sign-ins and credential additions, and look back to at least October 1, 2026. ARTEX operator targets the weaker supplier first Vendor's multi-tenant OAuth application stolen vendor credentials, or a new secret added to the existing app T1098.001 · T1199 Customer tenant A Customer tenant B Customer tenant C Customer tenant D …2,250+ tenants already consented in each tenant, so Graph API calls look like the vendor DEFENDER CHECK, IN EVERY TENANT Search by application ID, not display name · review service principal sign-ins, successes and failures separately · audit credential additions and consent grants · look back to at least October 1, 2026
Trusted relationship abuse at scale. The attacker never phishes the customers; it inherits the access they already granted their vendor. (The FLASH lists the pieces; this is how they fit together.)

ARTEX TTPs mapped to MITRE ATT&CK

The FLASH maps observed ARTEX activity across the whole intrusion lifecycle. Here is the condensed version. Technique IDs link to our interactive ATT&CK matrix.

Phase What ARTEX agents did Key ATT&CK techniques
Reconnaissance Mapped the effective permissions of stolen Azure credentials (SAS tokens, service principals, Kudu, storage keys) with sub-second test writes and ARM dry-runs; tested AWS IAM users with DryRun flags; fuzzed 500+ web paths; looked up victim geolocation T1595, T1580, T1526, T1087.004
Initial access Unsigned SAML assertions, XXE, SSRF, SSTI and command injection, SQL injection, IDOR, missing API authentication, admin passwords hardcoded in single-page-app JavaScript, stolen VPN configs, partner-tenant trust and vendor B2B federation T1190, T1078, T1199, T1195, T1606.002, T1552.001
Foothold and persistence JSP and PHP web shells (including via MySQL log poisoning), commands through Kudu admin APIs, open-source RATs (WISP, vshell, AdaptixC2), new SSH keys and Elastic IPs, password-reset account takeover, new credentials on existing OAuth apps T1505.003, T1098.001, T1078.004, T1105
Privilege escalation Launched instances with high-value IAM roles, Kerberoasting, DCSync of the krbtgt key, an is_admin=true flag on new accounts, resource-based constrained delegation, cloud metadata via XXE, admin tokens forged with unrotated keys T1558.003, T1098.003, T1552.005, T1606
Internal recon Iterative PowerShell against domain controllers, 110,000+ AD objects enumerated, LAPS passwords read, inventory of which hosts lacked an EDR sensor T1087.002, T1518.001, T1059.001
Lateral movement Pass-the-Ticket, Pass-the-Hash, WinRM and Kerberos relay toward DCs, blocked by endpoint protection T1550.003, T1550.002, T1021.006
Collection and exfil prep Backend SQL queries, mass blob reads, EC2 boot scripts to harvest data, mailbox reads, output sent to rotating webhook.site URLs, new 100 GB volumes for staging T1530, T1114.002, T1074
Command and control Residential proxies, Tor and SOCKS chains, Chisel and SoftEther tunnels, the BLOODALCHEMY backdoor T1090, T1572, T1071.001

Two details in that table matter more than the rest. First, the lateral movement attempts toward domain controllers were blocked by endpoint protection, and during internal recon the agents inventoried sensor coverage to find Windows hosts without CrowdStrike Falcon. EDR still works, and gaps in it are now something attackers enumerate automatically. Second, almost every initial access technique is a web application or identity flaw that a code review, a secrets scanner or a SAML configuration check would catch. ARTEX is not beating good controls. It is finding the places where controls were never applied.

The ARTEX intrusion chain in six phases with the control that breaks each one. Reconnaissance: fuzzes web paths and tests stolen cloud keys; break it with a full inventory of exposed apps. Initial access: SSRF, XXE, IDOR, unsigned SAML, secrets in JavaScript; break it with server-side authorization, signed SAML and secret scanning. Foothold: web shells, Kudu commands, new OAuth app secrets; break it by taking Kudu off the internet and alerting on new app credentials. Privilege escalation: Kerberoasting, DCSync, metadata theft, forged tokens; break it by rotating keys and enforcing IMDSv2. Internal recon and lateral movement: AD enumeration and hunting for hosts without EDR; break it with full EDR coverage. Collection: SQL dumps, mass reads, staging volumes; break it with volume alerts. A bracket marks the first four phases as under three hours in some intrusions. PHASE WHAT THE AGENTS DID WHAT BREAKS IT UNDER 3 HOURS Reconnaissance T1595 · T1580 fuzzes 500+ web paths; tests stolen cloud keys with sub-second checks Inventory every exposed app, including the forgotten ones Initial access T1190 · T1199 · T1606.002 SSRF, XXE, IDOR, unsigned SAML, secrets in JS bundles, partner trust Server-side authorization, signed SAML only, secret scanning Foothold T1505.003 · T1098.001 web shells, Kudu API commands, new OAuth app secrets, RATs Kudu/SCM off the internet; alert on new app credentials Privilege escalation T1558.003 · T1552.005 Kerberoasting, DCSync, metadata via XXE, tokens forged from old keys Rotate signing keys, enforce IMDSv2, long service passwords Recon and lateral T1087.002 · T1518.001 reads 110,000+ AD objects and LAPS; finds hosts without EDR EDR on every host (it blocked the lateral movement attempts) Collection T1530 · T1114.002 · T1074 dumps SQL, mass blob and mailbox reads, stages 100 GB volumes Alert on read volume and new large volumes or snapshots Command and control (proxies, Tor, Chisel, SoftEther tunnels) runs underneath every phase.
Every phase is a place to break the chain, and the controls on the right are not new. What ARTEX removes is the time you used to have between phases to notice.

ARTEX indicators of compromise

The FBI is explicit that these are historically observed infrastructure, much of it on cloud or VPN hosts that may since have been reassigned. Corroborate any hit with your own telemetry before blocking or declaring an incident. Indicators are defanged.

Indicator Context (per the FLASH)
38.244.50[.]120 ARTEX host and open directory, South Korea financial campaign
144.172.64[.]60, 144.172.64[.]99, 144.172.64[.]128 ConnectWise campaign: sign-ins and Graph API activity in customer tenants
187.15.88[.]186, 187.15.90[.]217, 187.15.89[.]186 Graph API access via compromised apps (transient VPN IPs, October 15–16, 2026)
101.53.80[.]20, 205.214.59[.]31, 124.155.252[.]63, 154.201.79[.]246, 23.248.249[.]90, 23.158.220[.]98, 103.248.148[.]84, 203.160.133[.]172, 209.209.85[.]38 Proxy IPs, South Korea financial campaign
199.19.95[.]66, 199.19.95[.]67, 199.19.95[.]68 WISP C2 nodes, Storm-3358 campaign (active as of November 6, 2026)
91.92.240[.]238, 85.11.168[.]45, 178.16.54[.]129, 31.77.229[.]35 ARTEX servers, Storm-3358 campaign (active as of November 6, 2026)
85.11.167[.]52 C2 server, Storm-3358 campaign (active as of November 6, 2026)
xcai[.]pro LLM API reseller used in the South Korea financial campaign

The behavioral indicators will outlive every IP on that list:

  • Outbound calls to LLM APIs such as api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai from systems that have no business making them, investigated alongside other suspicious activity.
  • High-frequency automated requests with sequential enumeration against login and search endpoints.
  • Outbound HTTPS transfers of structured personal data (names, phone numbers, financial details) to unapproved destinations.
  • .aspx or .jsp files in unexpected web application paths.
  • The fingerprints of an agent swarm: compressed bursts of activity, high volumes of failed attempts with small iterative changes between them, and duplicate or conflicting requests where sub-agents overwrite each other's work.

How to detect ARTEX activity

Agentic tooling is fast, but it is not quiet, and that is the defender's opening. The FLASH's "information requested" list doubles as a detection plan. Three places to look first:

1. Identity: new credentials on applications and service principals. Adding a secret or certificate to an app that already has broad consent is the quietest persistence in the ConnectWise campaign. In Microsoft Sentinel, start here and review every row:

AuditLogs
| where TimeGenerated > datetime(2026-10-01)
| where OperationName == "Add service principal credentials"
     or OperationName has "Certificates and secrets management"
     or OperationName has "Consent to application"
| extend Actor = coalesce(tostring(InitiatedBy.user.userPrincipalName),
                          tostring(InitiatedBy.app.displayName))
| extend Target = tostring(TargetResources[0].displayName),
         ObjectId = tostring(TargetResources[0].id)
| project TimeGenerated, OperationName, Actor, Target, ObjectId, Result
| order by TimeGenerated desc

Pair it with the service principal sign-in logs, and as the FBI advises, review successful and failed sign-ins separately and look beyond the published IPs at any source outside the vendor's expected infrastructure.

2. Egress: servers talking to LLMs. A web server, database host or build agent calling a model API is worth a question even when it turns out to be legitimate. In Microsoft Defender XDR:

DeviceNetworkEvents
| where Timestamp > ago(30d)
| where RemoteUrl has_any ("api.deepseek.com", "open.bigmodel.cn",
                           "api.x.ai", "api.anthropic.com", "xcai.pro")
| summarize Connections = count(), FirstSeen = min(Timestamp),
            LastSeen = max(Timestamp)
         by DeviceName, InitiatingProcessFileName, RemoteUrl
| order by Connections desc

Build an allow list of the hosts and processes that are supposed to use AI APIs. Everything else, and any unfamiliar LLM proxy, reseller, or self-hosted gateway such as CLIProxyAPI or New API in your egress logs, deserves a look.

3. Cloud control plane: the permission-mapping burst. ARTEX tests stolen cloud credentials in a pattern humans do not produce: test blob writes and deletes with sub-second timing, Azure ARM deployment validation dry-runs across many resource types, AWS calls with the DryRun flag, unexpected role assumptions and PassRole operations, and unusual Kudu/SCM API activity. Alert on the burst, not the individual call.

How to defend against ARTEX: the FBI's mitigations

The FLASH groups its recommendations into four areas. Condensed, with the reasoning behind each:

Reduce the external attack surface. ARTEX wins on the systems you forgot, so the first job is knowing what you expose.

  • Inventory every internet-facing application, explicitly including partner and supplier portals, mobile backends, employee and sales-support apps, legacy services, third-party integrations and anything a business unit built on its own.
  • Put cloud administration, network device management, Kudu/SCM publishing endpoints, hypervisors and backup platforms behind trusted networks, private connectivity or dedicated admin paths.
  • Disable basic publishing authentication where it is not needed, restrict who can read deployment credentials, and consider run-from-package deployments so deployed files cannot be modified in place.
  • Require an architecture review before a workload or management interface gets a public IP, and move enterprise apps behind Zero Trust access where you can.

Harden identity and access. The supply chain campaign ran through app consent, not phishing.

  • Apply least privilege to multi-tenant app permissions and consent grants, and monitor credential additions, privilege assignments and app-only activity on highly privileged apps.
  • If you ever installed an affected ConnectWise application, review every Microsoft 365 tenant where it is or was authorized, even if you stopped using it, and search by application ID, not display name.
  • Audit Entra for new apps, service principals and owners, app-role assignments, consent grants, privileged role assignments and changes to Conditional Access. If something is confirmed, revoke and rotate.
  • Require approval for higher-risk certificate enrollment, enable Extended Protection for Authentication on CA web interfaces, and watch for unusual enrollment by privileged users.

Strengthen application protections. Machine-speed enumeration is only a problem if your applications let it run.

  • Put rate limits, recipient limits and transaction thresholds on high-impact business APIs, and alert when an application runs far outside its normal volume.
  • On partner and employee portals, alert on sequential lookups and unusually high records-per-session, enforce server-side authorization on every customer record, and return only the fields the operation needs.
  • Stop returning stack traces, SQL errors, internal paths and framework details to unauthenticated users. Every error message is a hint to an agent that iterates on feedback.
  • Scan repositories, container images, build artifacts, pipelines, environment variables and infrastructure-as-code for embedded secrets continuously, and rotate anything exposed. Hardcoded admin passwords in front-end JavaScript were an initial access vector.

Preserve evidence and respond. Speed cuts both ways: do not let evidence handling slow containment.

  • Preserve Entra sign-in and audit logs, Purview Audit records, SIEM data and Graph telemetry, but do not delay urgent containment while exports run.
  • For ConnectWise-related activity, review from at least October 1, 2026, including failed authentications and volume anomalies.
  • Have an authorized responder remove confirmed unauthorized credentials, grants, owners and identities only after evidence is preserved, and keep affected ConnectWise connections disabled until the vendor issues restoration instructions. Do not delete the enterprise application as a first step.

If you are working an incident, the FBI asks for ARTEX console screenshots or config files, sign-in and audit logs, LLM egress evidence, cloud activity logs, persistence artifacts and impact figures, reported to your local field office or IC3. Our incident response lifecycle post covers the evidence-first sequencing behind those instructions.

The uncomfortable part: ARTEX did nothing a skilled human team could not do. The FBI's own table shows endpoint protection stopping the lateral movement, and every initial access technique is a known, testable flaw. What changed is the economics. An attacker with little skill and a cheap API key can now probe every forgotten portal you own, in parallel, and keep going while your team sleeps. The answer is not an AI arms race. It is closing the basic gaps on the systems you were not watching, and alerting on the shapes of automated activity so a human gets the page before the agent finishes.

What this means for your certification

Security+ (SY0-701 and SY0-801). ARTEX is a case study in reducing the attack surface, application attacks (SSRF, XXE, SQL injection, IDOR) and supply chain risk. The new SY0-801 exam (available from November 17, 2026) adds objective 2.6 on AI threats and vulnerabilities, and "an attacker uses an AI agent framework to automate an intrusion" is the kind of scenario it covers. Review our application attack indicators and third-party risk deep dives.

CySA+. The FLASH is a ready-made exercise: take each observed behavior and map it to an ATT&CK tactic, then write the detection. The TTP table above and our threat hunting post are a good place to practice.

SecAI+. This is AI-enabled attack automation in the wild, the material in Domain 3, and its flip side, why an agent's permissions decide its blast radius. Expect questions on how AI changes attacker speed and scale, and which controls still hold.

ARTEX FAQ

Is ARTEX malware?

No. ARTEX is a legitimate-looking, open-source penetration testing framework that orchestrates AI agents. It becomes an attack tool when someone points it at systems they are not authorized to test. That is why detection focuses on behavior and on the tools it deploys (web shells, RATs such as WISP, vshell and AdaptixC2) rather than on ARTEX itself.

Did taking down the GitHub repository stop it?

No. The developer removed it on October 8, 2026, but forks, localized variants, archived copies and the commercial ARTEX PRO remain available, and researchers found more than 1,600 exposed ARTEX consoles in early November.

Which AI models does ARTEX use?

Whatever the operator connects. The FBI lists Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. In attacker-controlled instances, Chinese models dominate, especially DeepSeek's Flash family, often accessed through API resellers.

Should we block all LLM API traffic?

Not blindly. Plenty of legitimate software calls these APIs. The FBI's guidance is to investigate unexpected calls alongside other suspicious activity. Allow-list the systems that should use AI services and alert on everything else, especially servers.

How fast is an ARTEX intrusion?

The FLASH cites intrusions that went from initial targeting to privileged access in under three hours. Plan your alerting and on-call escalation around that number.

3 takeaways you can act on today
  • Close the admin side doors. Find every Kudu/SCM publishing endpoint, backup console, hypervisor and cloud or network management interface reachable from the internet and move it behind a trusted network or private path. Turn off basic publishing authentication where nothing depends on it. These were direct footholds in the FLASH.
  • Audit your app credentials and consent grants back to October 1. Run the Entra audit query above, review every new secret, certificate, owner and consent grant on applications and service principals, search vendor apps by application ID, and turn the query into a standing alert.
  • Put a speed limit on your portals and watch your egress. Add rate limits and alerts for sequential lookups and high records-per-session on partner, customer and employee portals, and alert when a server that is not on your AI allow list calls an LLM API.

Source: FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework," November 13, 2026 (TLP:CLEAR).

#ARTEX #AgenticAI #ThreatIntelligence #FBI #MITREATTACK #IncidentResponse #SecAIplus #CySA #SecurityPlus #CyberSecurity #InfoSec

Domain 3 · AI-assisted security

Turn this into exam points

AI-assisted Security is 24% of SecAI+. Study it end to end, objective by objective, in the SecAI+ Collection.

  1. Find your gapsA free readiness quiz scores you domain by domain.
  2. Study what's on the testWork through every objective, with practice questions.
  3. Book it with confidenceSchedule the exam when you're scoring well in every domain.

Read a free excerpt from SecAI+ Domain 3: AI-assisted Security

Share this article

Keep reading

AI attacks & defenses

MITRE ATT&CK v20: What's New, and What It Means for Your Exam

Nine new Enterprise techniques, five of them about AI: prompt injection, AI-enabled exfiltration, attacks on AI-assisted defenses and poisoned AI search. Released October 13, 2026, here is every addition, where it sits in the matrix, the defenses that matter, and what changes for Security+, CySA+ and SecAI+ candidates.

Read
AI attacks & defenses

Normal Is a Moving Target: Baselines, UEBA, the Base-Rate Fallacy, and How AI Security Tools Decide Something Is Wrong

A 99% accurate detector sounds like a solved problem until you run the numbers on a network where one event in ten thousand is malicious: nearly every alert it raises is wrong. SecAI+ objective 3.1 tests how AI-enabled tools decide something is abnormal. Baselining and UEBA, anomaly versus signature detection, supervised versus unsupervised detection, precision and recall, the base-rate arithmetic worked with real numbers, model drift, and the human review loop that makes any of it usable.

Read
AI attacks & defenses

Deepfakes, Voice Clones, and Machine-Speed Phishing: How AI Rewrites the Attacker's Playbook

AI did not invent a single new attack. It made every existing one cheaper, faster and personal. SecAI+ objective 3.2 in full: synthetic media and impersonation, misinformation versus disinformation, automated reconnaissance and data correlation, AI-written payloads and obfuscation, and the verification ladder that stops a cloned voice from moving money.

Read

Follow Skillthropic

Keep learning between study sessions

The same exam-focused teaching, in smaller pieces, wherever you scroll.

Enjoyed this? Get the AI security news that matters.

Join The AI Security Brief for the top AI security news, plus what's important to the C-suite. Free, straight to your inbox.

No spam. Unsubscribe anytime.