ARTEX is an open-source penetration testing framework that hands the work of an attack to a team of AI agents. An operator types in a target and a goal, such as gaining admin access. A planner agent decides what to try next, worker agents run the tools, and every result is written to a shared graph that the next step reads from. Microsoft describes it as running a test “from reconnaissance through data exfiltration with little needed human interaction”. It was built for authorized testing. Within weeks of its release in July 2026 it was being used in real intrusions, first against South Korean financial institutions and then by Storm-0940, a China state-aligned actor, in the ConnectWise Cloud Backup supply chain compromise. Since then the same tooling has been seen against identity providers, software vendors and AWS accounts. ARTEX does not bring new techniques. It brings speed and persistence. Microsoft measured 2,900 failed sign-ins a minute from one Storm-0940 address, and cases where the actor went from first targeting to privileged cloud access in under three hours. The same automation makes it noisy and repetitive, which is what makes it detectable. This page explains how ARTEX works, how attackers use it, what it leaves in your logs and how to hunt for it. What ARTEX is Fact Detail Released Published on GitHub as Autumn-27/ARTEX on July 26, 2026, under the AGPL-3.0 license Developer Autumn-27, who also wrote ScopeSentry, a Chinese asset discovery and vulnerability scanning platform Status Made closed source on October 8, 2026, after it had passed 12,000 forks. The repository has been taken down. Archived copies, translations and renamed builds still circulate Built with Go, a Next.js web console and PostgreSQL. The console listens on port 8787 by default AI models Any model. It speaks the OpenAI Chat Completions format, the OpenAI Responses format and the Anthropic Messages format, which covers most commercial, reseller and self-hosted models Other names ScopeWeaver, LUCK, Guanlan SEC, ARTEX MX and ARTEX v2 consoles have been seen running the same platform How it runs an attack. These are the parts of the platform, as described by the researchers who examined it: Task. The operator defines the scope, any test accounts and the goal. Each task carries allow and deny rules, a timeout and a wrap-up step. Planner agent. Turns the goal into attack “intents” and hands them out. Worker agents. Claim intents and run tools such as nmap, curl and Impacket, three at a time by default, then write the results back. Graphs. An asset graph records what the target looks like. An exploration graph records what has been tested. Facts, findings and hints sit on a shared “blackboard” that every agent reads. Traffic recording. mitmproxy is included, so requests and responses are kept. Model failover. When one AI provider bans or cuts off a key, ARTEX moves to the next. A single campaign can be spread across several providers and keys. Recon feed. It syncs domains, subdomains, IP addresses, ports and endpoints found by ScopeSentry straight into its asset graph. Some versions ship with a default blocklist that excludes .gov, .gov.cn, .edu and .edu.cn targets. The operator can change it. The playbook is not fixed. Datadog notes that the builds it analyzed have no set method for cloud post-exploitation, so what the agents do after they get credentials depends on the AI model behind them. Hunt for the behavior, and expect the details to vary. How attackers use it When Who Targets What happened Late September to early October 2026 An unattributed actor. CrowdStrike assesses with moderate confidence a financially motivated Chinese speaker South Korean financial organizations A loan inquiry service used by brokers was breached at one bank and an employee mobile work system at another. Data was exfiltrated. The ARTEX instance used a DeepSeek model bought through a reseller From September 2026 Storm-0940, China state-aligned, active since 2021 Hundreds of organizations in aerospace and defense, energy, financial services, government, legal, software and other sectors ARTEX automated scanning, password spraying and OAuth application abuse. In the ConnectWise compromise it attempted sign-ins to more than 2,250 Microsoft 365 tenants October 2026 An ARTEX-using actor that overlaps with Storm-0940 IT and software vendors Sustained attacks on Okta, Entra ID and Amazon Cognito sign-in, on vendor web applications and on Microsoft 365 integrations, aiming at the vendors’ customer data October 2026 ARTEX operators observed by Datadog AWS accounts, with a focus on financial data Compromised long-term access keys were used to search CloudWatch logs for secrets and financial records. Public Cognito user pools were abused to self-register and collect AWS credentials Microsoft says it has identified threat actors across multiple regions using ARTEX, so treat it as shared tooling, not the mark of one group. Our ConnectWise Cloud Backup advisory covers the Storm-0940 supply chain compromise in detail. Exposed consoles. Operators run ARTEX as a hosted web console, and many leave it facing the internet. Censys counted 1,465 hosts with an exposed ARTEX console in early November 2026, most in the United States, Hong Kong and China. About 650 of them sit in two blocks of identically configured hosting, which points to a provisioning template. Some consoles belong to authorized testers. Others share a host with Cobalt Strike, Vshell, Adaptix C2 or other AI attack frameworks such as CyberStrikeAI and PentAGI. What an ARTEX attack looks like Wiz, Microsoft and Datadog watched the agents work in customer logs. The steps repeat across victims. Read the application. Minutes after first contact, the agent downloads JavaScript bundles, source maps, runtime configuration and API documentation: /packs/js/main-<hash>.js.map, /assets/config.json, /env.js, /swagger/?format=openapi. Soon after, it calls internal API routes by name. Those names appear only in the source maps it just read. Learn what “not found” looks like. Before each round of guessing, it requests a path that should not exist, such as /zzz-nonexistent-xyz123. Against S3 the same habit shows as object keys like zzz_nonexistent_abc.txt and Notexist-xyz. Probe sign-in. It manipulates OAuth callbacks with made-up codes and broken state values, and tries unsigned tokens. Wiz saw one forged token, which decodes to {"alg":"none"} and {"email":"test@test.com"}, sent to more than 60 variations of an SSO route in two minutes. Spray passwords. Against Okta and Entra ID, in bursts of 20 to 60 minutes, with a median gap of one second between attempts and a new user agent for each burst. Software engineers were the main targets, and some accounts got more than a dozen attempts. Abuse account recovery and sign-up. It checks which usernames exist, then hammers reset and activation endpoints. Where sign-up is open, it registers with a disposable mailbox, confirms the account and immediately asks for another tenant’s data. Use what it gets. With stolen OAuth application credentials, Storm-0940’s agents signed in as the application, explored Microsoft Graph, added credentials to applications for persistence and tried to change identities, groups and privileged roles. With a stolen AWS access key, the agent searched CloudWatch logs for secrets, file links and internal addresses. Enumerate everything it can reach. In AWS it tries nearly every read-only discovery call. In Azure, Microsoft saw bulk secret reads across several Key Vaults. Adapt and keep going. When a login page starts rate limiting, the agent moves to another endpoint on the same host. When an address is blocked, a new one appears within minutes. Some sessions ran for 24 to 30 hours. What gives it away Behavior Where it shows Why it stands out Sign-in attempts exactly one second apart, for 20 to 60 minutes Entra ID and Okta sign-in logs People and most spray tools do not hold a fixed one-second rhythm for an hour A request for a placeholder path right before a burst of guesses Web server, WAF and S3 access logs /zzz-nonexistent-xyz123 and similar names have no legitimate caller Internal API routes called minutes after source maps are downloaded Web server and WAF logs The route names are not published anywhere else The same unsigned token sent to dozens of SSO paths Web server and WAF logs The token begins eyJhbGciOiJub25lIn0. which is alg: none User agents that name the tool Sign-in, Graph, CloudTrail and web logs Operators do not always change the defaults, such as artex-pentest/1.0 FilterLogEvents always called with a limit of 1000 AWS CloudTrail The constant repeats across sessions, with search terms like ?password ?AKIA A sweep of read-only API calls that all fail AWS CloudTrail A new identity trying every discovery call produces a spike of access denied errors An application adding credentials or assigning itself a role Entra ID audit log Applications rarely change their own access Indicators of compromise A match needs investigation. No match proves nothing: addresses rotate within minutes, user agents are easy to change, and several listed below are common in normal traffic. User agents that name ARTEX. These are high-confidence indicators. artex-readonly/1.0 artex-pentest/1.0 artex-audit/1.0 artex-verify/1.0 artex-spawn/1 ARTEX-worker/1.0 artex-i21861-canary Mozilla/5.0 ARTeX Mozilla/5.0 (Windows NT 10.0; Win64; x64) ARTeX-authorized-pentest Mozilla/5.0 (X11; Linux x86_64) artex-verify/1.0 Other user agents seen in ARTEX activity. The first three are distinctive. The rest are common and only useful together with other signals. ReconProbe/1.0, norma/0.4, azure-toolkit/1.0 Python-urllib/3.12, python-requests/2.31.0, python-requests/2.34.2, curl/7.88.1 Truncated browser strings such as Mozilla/5.0, Mozilla/5.0 (Windows NT 10.0; Win64; x64) and Mozilla/5.0 (Windows NT 10.0; Win64; x64) Chrome/120 Boto3 1.43.112 and 1.43.113 and aws-cli 1.46.1 on os/linux#6.8.0-147-generic with lang/python#3.12.14 Request artifacts. Artifact Where to look /zzz-nonexistent-xyz123 Web and WAF logs zzz_nonexistent_abc.txt, Notexist-xyz S3 access logs and CloudTrail data events eyJhbGciOiJub25lIn0.eyJlbWFpbCI6InRlc3RAdGVzdC5jb20ifQ. jwt parameters and Authorization headers on SSO routes callback?code=AAAA&state=notvalidb64!! OAuth callback routes A tenant domain like artex-hqheizfk[.]invalid in sign-up attributes Cognito PreSignUp errors An attempt to set custom:isAdmin at sign-up Cognito SignUp errors Source addresses. Address Reported by Context 144.172.64[.]128, 144.172.64[.]99, 144.172.64[.]60 Microsoft Storm-0940 sign-ins, password spray and Graph activity 187.15.88[.]186, 187.15.90[.]217, 187.15.89[.]186 Microsoft VPN addresses, relevant to October 15 and 16 only 104.194.150[.]21, 167.88.172[.]97 Wiz Attacks on IT and software vendors 154.64.252[.]77, 154.36.136[.]197, 154.36.136[.]23 Datadog NetLab Global hosting. The first ran the Cognito sign-up spike of October 26 and 27 103.54.154[.]40, 103.54.154[.]32 Datadog Zeofast Network hosting. The first searched CloudWatch logs on October 20 169.128.55[.]214, 82.158.88[.]233, 45.145.228[.]25 Datadog Hosting providers 102.216.154[.]102, 88.229.22[.]232 Datadog ISP addresses in Kenya and Turkey 38.244.50[.]120 CrowdStrike ARTEX host used against South Korean financial organizations 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 CrowdStrike Proxy addresses used in the same campaign ARTEX consoles. Useful for attack surface checks and for egress logs. Page titles starting with ARTEX and a space, colon or dash, on a page that loads _next/static assets. Renamed consoles use ScopeWeaver, LUCK and 观澜SEC渗透测试平台. Default port 8787, then 8790, 443 and 80. API paths /tasks, /exploration/findings, /exploration/graph, /exploration/intents, /llm/profiles, /llm/pool, /intercept. xcai[.]pro, an AI model reseller used by the actor in the South Korean campaign. Censys also found the self-hosted model gateways CLIProxyAPI and New API on ARTEX hosts. Unfamiliar AI proxies and resellers in your egress logs are worth a look. Free SOC trialWant this watched across your environment without building the hunts yourself? Our 24x7 SOC monitors your endpoints, servers and cloud during a free trial.Start a free trialThreat hunting Hunt from September 1, 2026. Microsoft dates Storm-0940’s use of ARTEX to that month. The KQL below is example code for Microsoft Sentinel. Adapt table and field names to your environment and test each query before relying on it. The web log hunts use CommonSecurityLog. If your WAF, proxy or web server logs land in another table, swap the table and the URL and user agent columns. Hunt 1: indicator sweep Search every log source that records a user agent or a source address. // Example: ARTEX user agents and published addresses across identity, Graph, AWS and web logs let ArtexIPs = dynamic(["144.172.64.128","144.172.64.99","144.172.64.60","104.194.150.21","167.88.172.97", "154.64.252.77","154.36.136.197","154.36.136.23","103.54.154.40","103.54.154.32","169.128.55.214", "82.158.88.233","45.145.228.25","102.216.154.102","88.229.22.232","38.244.50.120","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"]); union isfuzzy=true (SigninLogs | project TimeGenerated, Source="EntraSignIn", Agent=UserAgent, IP=IPAddress, Detail=UserPrincipalName), (AADNonInteractiveUserSignInLogs | project TimeGenerated, Source="EntraNonInteractive", Agent=UserAgent, IP=IPAddress, Detail=UserPrincipalName), (MicrosoftGraphActivityLogs | project TimeGenerated, Source="Graph", Agent=UserAgent, IP=IPAddress, Detail=RequestUri), (AWSCloudTrail | project TimeGenerated, Source="CloudTrail", Agent=UserAgent, IP=SourceIpAddress, Detail=EventName), (CommonSecurityLog | project TimeGenerated, Source="WebOrProxy", Agent=RequestClientApplication, IP=SourceIP, Detail=RequestURL) | where TimeGenerated >= datetime(2026-09-01) | where Agent has "artex" or Agent has "ReconProbe" or Agent startswith "norma/" or Agent startswith "azure-toolkit/" or IP in (ArtexIPs) | summarize Events=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), Samples=make_set(Detail, 10) by Source, Agent, IP | order by FirstSeen asc Hunt 2: password spray at a fixed rhythm Find sources that fail sign-ins for many users with a steady gap of a second or two. // Example: failed sign-ins from one source and user agent, many users, median gap of two seconds or less union isfuzzy=true SigninLogs, AADNonInteractiveUserSignInLogs | where TimeGenerated >= datetime(2026-09-01) | where ResultType != "0" | project TimeGenerated, IPAddress, UserAgent, UserPrincipalName, ResultType | sort by IPAddress asc, UserAgent asc, TimeGenerated asc | extend GapSec = iff(IPAddress == prev(IPAddress) and UserAgent == prev(UserAgent), datetime_diff('second', TimeGenerated, prev(TimeGenerated)), long(null)) | summarize Attempts=count(), Users=dcount(UserPrincipalName), MedianGapSec=percentile(GapSec, 50), Lockouts=countif(ResultType == "50053"), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated) by IPAddress, UserAgent, bin(TimeGenerated, 1h) | where Attempts >= 200 and Users >= 20 and MedianGapSec <= 2 | order by Attempts desc In Okta, run the same check on the System Log: failed user.session.start events and user.account.lock events from one address across many users, at about one attempt a second. Hunt 3: a spray that ends in a successful sign-in // Example: a successful sign-in from an address that failed against many users let Sprayers = union isfuzzy=true SigninLogs, AADNonInteractiveUserSignInLogs | where TimeGenerated >= datetime(2026-09-01) and ResultType != "0" | summarize Failures=count(), TargetedUsers=dcount(UserPrincipalName) by IPAddress | where Failures >= 100 and TargetedUsers >= 20; union isfuzzy=true SigninLogs, AADNonInteractiveUserSignInLogs | where TimeGenerated >= datetime(2026-09-01) and ResultType == "0" | join kind=inner Sprayers on IPAddress | project TimeGenerated, UserPrincipalName, IPAddress, UserAgent, AppDisplayName, Failures, TargetedUsers | order by TimeGenerated asc Hunt 4: application reconnaissance in web logs Two or more of these markers from one source against one host is a strong sign of an agent reading and probing the application. // Example: source maps, runtime config, the not-found canary, forged tokens and recovery abuse CommonSecurityLog | where TimeGenerated >= datetime(2026-09-01) | extend Marker = case( RequestURL has "zzz-nonexistent" or RequestURL has "zzz_nonexistent" or RequestURL has "Notexist-xyz", "not-found canary", RequestURL has "eyJhbGciOiJub25lIn0", "unsigned token", RequestURL has "code=AAAA" or RequestURL has "notvalidb64", "forged OAuth callback", RequestURL endswith ".js.map", "source map", RequestURL has "/env.js" or RequestURL has "/assets/config.json", "runtime config", RequestURL has "format=openapi" or RequestURL has "swagger", "API documentation", RequestURL has "GeneratePasswordResetToken" or RequestURL has "ValidateUserName" or RequestURL has "forgot_password", "account recovery", "") | where isnotempty(Marker) | summarize Requests=count(), Markers=make_set(Marker), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), Agents=make_set(RequestClientApplication, 10) by SourceIP, DestinationHostName | where array_length(Markers) >= 2 or Markers has "not-found canary" or Markers has "unsigned token" | order by Requests desc Hunt 5: applications changing their own access Storm-0940’s agents added credentials to applications and one application assigned itself Global Administrator. // Example: credential additions and role or permission grants made by an application, not a person AuditLogs | where TimeGenerated >= datetime(2026-09-01) | where OperationName in ("Add service principal credentials", "Add member to role", "Add app role assignment to service principal") or OperationName has "Certificates and secrets management" | extend ActorApp = tostring(InitiatedBy.app.displayName), ActorSpId = tostring(InitiatedBy.app.servicePrincipalId) | where isnotempty(ActorSpId) | extend Target = tostring(TargetResources[0].displayName), TargetId = tostring(TargetResources[0].id) | extend SelfChange = ActorSpId == TargetId | project TimeGenerated, OperationName, ActorApp, ActorSpId, Target, TargetId, SelfChange, Result | order by TimeGenerated asc Check every row against a change record. Automation accounts will appear here, so build a list of the ones you expect. Hunt 6: scripted Microsoft Graph bursts by an application This needs Microsoft Graph activity logs sent to your workspace. Graph reads are not in the Entra audit log. // Example: service principals calling Graph with scripted clients, at volume or with many denials MicrosoftGraphActivityLogs | where TimeGenerated >= datetime(2026-09-01) | where isnotempty(ServicePrincipalId) | where UserAgent has "urllib" or UserAgent has "python-requests" or UserAgent startswith "curl/" or UserAgent startswith "azure-toolkit/" or UserAgent has "artex" | summarize Requests=count(), Denied=countif(ResponseStatusCode in (401, 403)), Endpoints=dcount(tostring(split(RequestUri, "?")[0])), IPs=make_set(IPAddress, 10) by AppId, ServicePrincipalId, UserAgent, bin(TimeGenerated, 1h) | where Requests >= 100 or Denied >= 20 | order by Requests desc Hunt 7: bulk Key Vault secret reads // Example: one identity reading many secrets across several vaults in 15 minutes AzureDiagnostics | where TimeGenerated >= datetime(2026-09-01) | where ResourceProvider == "MICROSOFT.KEYVAULT" and OperationName == "SecretGet" | summarize SecretGets=count(), Vaults=dcount(Resource), IPs=make_set(CallerIPAddress, 10) by identity_claim_appid_g, bin(TimeGenerated, 15m) | where SecretGets >= 20 and Vaults >= 3 | order by SecretGets desc Hunt 8: CloudWatch log searches with a stolen AWS key // Example: FilterLogEvents with the fixed limit of 1000 or secret-hunting search terms AWSCloudTrail | where TimeGenerated >= datetime(2026-09-01) | where EventName == "FilterLogEvents" | extend Params = parse_json(RequestParameters) | extend Limit = toint(Params.limit), Pattern = tostring(Params.filterPattern), LogGroup = tostring(Params.logGroupName) | where Limit == 1000 or Pattern has_any ("AKIA", "password", "presigned", "mongo") or Pattern has "s3://" | summarize Calls=count(), LogGroups=dcount(LogGroup), Patterns=make_set(Pattern, 50) by UserIdentityArn, UserIdentityAccessKeyId, SourceIpAddress, bin(TimeGenerated, 1h) | where Calls >= 20 | order by Calls desc Read the search patterns in the result. Company names combined with words like Fund, Invoice, Billing and LLC match what Datadog saw. Hunt 9: Cognito sign-up abuse and the discovery sweep after it // Example: bursts of Cognito sign-up and recovery calls from one source AWSCloudTrail | where TimeGenerated >= datetime(2026-09-01) | where EventSource == "cognito-idp.amazonaws.com" | where EventName in ("SignUp", "InitiateAuth", "ForgotPassword", "ResendConfirmationCode", "RespondToAuthChallenge") | summarize Calls=count(), Errors=countif(isnotempty(ErrorCode)), Events=make_set(EventName), Agents=make_set(UserAgent, 10) by SourceIpAddress, bin(TimeGenerated, 1h) | where Calls >= 30 | order by Calls desc // Example: a Cognito-issued identity trying many discovery calls that are denied AWSCloudTrail | where TimeGenerated >= datetime(2026-09-01) | where UserIdentityArn has "CognitoIdentityCredentials" | where ErrorCode has "AccessDenied" or ErrorCode has "Unauthorized" | summarize Denied=count(), Actions=dcount(EventName), Sample=make_set(EventName, 25) by UserIdentityArn, SourceIpAddress, bin(TimeGenerated, 1h) | where Actions >= 8 | order by Denied desc Every ARTEX session Datadog reviewed opened with unauthenticated S3 reads. In CloudTrail data events these have a userIdentity.type of AWSAccount and an accountId of anonymous. Check S3 data events or server access logs for anonymous GetObject and ListObjects calls, and for the placeholder object names listed above. Detection rules to keep These hold up after the published addresses and user agents change. Failed sign-ins for many users from one source at a steady rate, in Entra ID and Okta. Many failed sign-ins across users followed by a successful sign-in from the same source, above all to an admin account. A credential added to an application or service principal by another application. A service principal assigning a directory role or app role to itself. An application calling Microsoft Graph from a network it has never used. Bulk secret reads across several Key Vaults by one identity. An AWS identity with an unusual number of access denied errors in a short window. Read-only discovery calls across many AWS services by a long-term access key or a new identity. A Cognito sign-up burst from one source, and Cognito-issued credentials used for discovery calls. Requests for source maps or runtime configuration followed by calls to undocumented API routes. Any request carrying an alg: none token. If you find ARTEX activity Preserve the logs first. Export sign-in, audit, Graph, CloudTrail and web logs for the period. Entra sign-in records age out in 30 days on P1 and P2 tenants. Work out how far it got. Failed attempts are reconnaissance. A successful sign-in, an issued token or a first successful API call is the point where it becomes an incident. Identity. Reset the affected accounts, revoke their sessions and review the MFA methods registered on them. For applications, remove credentials you cannot account for, then review role assignments, consent grants and owners. AWS. Deactivate and rotate the exposed access key. Review everything that key did, starting with CloudWatch log searches, Secrets Manager and S3. Treat any secret that appeared in a searched log group as exposed. Web applications. Remove accounts the agent registered. Check whether any cross-tenant request returned data. Fix the route that answered. Block, then keep watching. Block the source addresses and user agents, and expect new ones within minutes. Keep the behavioral rules above running. Closing the gaps it looks for Take source maps and runtime configuration off production. They hand the agent your internal API routes. Reject unsigned tokens and validate OAuth state on every callback. Rate limit by account as well as by address, and apply the limit across every sign-in and recovery endpoint on a host. Require phishing-resistant MFA for administrators and engineers, the accounts the sprays focused on. Close open self-registration where you do not need it. On Cognito, validate sign-up attributes, block client-writable admin attributes and give the identity pool’s authenticated role the least access that works. Rotate or remove long-term AWS access keys, and keep secrets out of logs. Apply least privilege to multi-tenant applications and review vendor applications that hold admin consent, including ones you no longer use. Send Microsoft Graph activity logs and S3 data events to your SIEM. Without them the most useful evidence is not recorded. Timeline July 26, 2026: ARTEX is published on GitHub. September 2026: ARTEX wins Baidu’s Agent+ red team challenge. Microsoft dates Storm-0940’s use of it to as early as this month. Late September to early October 2026: The campaign against South Korean financial organizations. October 7, 2026: CrowdStrike reports that campaign. October 8, 2026: The developer makes ARTEX closed source and says the project was meant for authorized testing. October 10 to 14, 2026: Storm-0940 sprays 327 accounts in eight organizations, with more than 151,000 failed sign-ins and no success. October 15 and 16, 2026: Storm-0940’s agents sign in to Microsoft 365 tenants through compromised ConnectWise applications. October 20, 2026: A stolen AWS key is used to search CloudWatch logs for secrets. October 21, 2026: Microsoft attributes the ConnectWise compromise to Storm-0940 and ARTEX. October 22, 2026: Wiz reports the same tooling against IT and software vendors. October 26 and 27, 2026: A Cognito sign-up spike from 154.64.252[.]77. November 5, 2026: Censys reports 1,465 exposed ARTEX consoles. November 9, 2026: Datadog reports ARTEX activity in AWS environments. Frequently Asked Questions Is ARTEX malware? No. ARTEX is a penetration testing framework that runs on the operator’s own server and attacks targets over the network. Nothing is installed on the victim by ARTEX itself, so endpoint antivirus will not find it. The evidence is in sign-in logs, cloud audit logs and web server logs, where the agents’ requests are recorded. Does ARTEX have a CVE or a patch? No. ARTEX is a tool, not a vulnerability, so there is no CVE and nothing to patch. It finds and uses weaknesses that already exist: weak passwords, exposed source maps, open sign-up, stolen application credentials and long-lived access keys. The fix is to close those gaps and to detect the behavior described on this page. Who is using ARTEX? Microsoft has tied ARTEX to Storm-0940, a China state-aligned actor active since 2021, and says it has seen threat actors in multiple regions using the framework. CrowdStrike reported a separate, unattributed actor who used it against South Korean financial organizations and is assessed as likely financially motivated. Treat ARTEX as shared tooling. Can I block ARTEX by its user agent? Partly. Some operators leave the default user agents in place, such as artex-pentest/1.0, and blocking those is a quick win. The user agent is easy to change, and Wiz saw more than 15 different strings across victims, with a new one for each password spray burst. Use user agents for hunting and rely on behavioral rules for detection. Is every exposed ARTEX console malicious? No. Censys notes that some of the consoles it found are likely run by authorized testers. The console does not need to face the internet at all, so exposure says more about the operator’s care than intent. A console on a host that also runs Cobalt Strike or Vshell, or one seen in your own logs, deserves attention. Which AI models does ARTEX use? Any the operator chooses. ARTEX talks to OpenAI-compatible and Anthropic-compatible endpoints, which covers commercial, reseller and self-hosted models. The actor in the South Korean campaign ran it on a DeepSeek model bought through a reseller. When one provider cuts off a key, ARTEX switches to the next, so blocking one provider does not stop a campaign. How fast does an ARTEX attack move? Fast enough that alerts alone are not a defense. Microsoft recorded 2,900 failed sign-ins a minute from one Storm-0940 address, and multiple cases where the time from first targeting to privileged cloud access was under three hours. The activity generates plenty of alerts. The problem is the time left to act on them. How can I tell an AI agent from a human attacker in the logs? Look for timing and chaining. Agents hold a fixed rhythm, such as one request a second, for an hour. They use the output of one step in the next within minutes, for example calling routes found in a source map they just downloaded. They also keep testing things an experienced person would skip, and run for a day or more without a pause. How ThreatDefence is responding ThreatDefence has added the published ARTEX addresses, user agents and request artifacts to our threat intelligence. Across all monitored environments, our SOC is: sweeping sign-in, cloud and web logs for the published indicators, back to September 1 reviewing application credential additions and role changes in Entra ID watching for fixed-rhythm password sprays and for enumeration bursts in Microsoft 365, Azure and AWS If you want help running a retrospective hunt or reviewing what an application or access key did, contact us. Read more about our AI SOC platform, identity threat detection and response and 24x7 SOC. Sources Microsoft Security Blog, AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise (October 21, 2026) Datadog Security Labs, On the hunt for ARTEX: understanding agentic attacks in AWS environments (November 9, 2026) Wiz, ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software Vendors (October 22, 2026) Censys, ARTEX: Agentic Asbestos (November 5, 2026) CrowdStrike, Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance (October 7, 2026)