DEV Community

Anoymask
Anoymask

Posted on Edited on

FBI ARTEX Threat Overview: TTPs, IOCs, and Investigation Items for Multi-Agent Attacks

1. Overview

2. Executive Summary

The FBI synthesized cases where multiple threat actors used ARTEX to automate reconnaissance, authentication and cloud privilege verification, vulnerability exploitation, persistence, lateral movement, and data staging. In the Barracuda incident, a specific attack chain was confirmed where credential chaining moved from a cloud key left in a third-party vendor's GitHub configuration to roughly 10 cloud accounts, analytics, source repositories, and Kubernetes within approximately 44 hours.

3. Attack Flow

Typical Flow Observed Across Multiple Campaigns

The following summarizes observations compiled by the FBI across multiple campaigns and does not imply that the exact same order or techniques were used in every incident.

  1. ARTEX agents perform WHOIS lookups, directory fuzzing, and vulnerability scans while rapidly checking the effective permissions of cloud credentials via dry-runs or test writes.
  2. The framework iteratively explores reachable attack paths, such as public application vulnerabilities, weak credentials, leaked VPN configurations, trust relationships with external partner tenants, and unsigned SAML assertions.
  3. On successful paths, it attempts to establish footholds and persistence using web shells, the Kudu administrative API, open-source RATs, additional cloud credentials, SSH keys, or OAuth application credentials.
  4. It enumerates domain controllers and cloud resources, attempting lateral movement via Pass-the-Hash/Pass-the-Ticket, WinRM, or Kerberos relay. FBI documentation also includes lateral movement attempts blocked by endpoint protection.
  5. It explores mailboxes, cloud storage, and backend databases, staging data using volumes or scripts to prepare for exfiltration.

Credential Chaining Observed in Barracuda

  1. A compromise of a third-party AI vendor's development environment yields a long-lived Barracuda cloud access key left in a GitHub configuration file.
  2. The key is used to explore secret stores, infrastructure-as-code state, server metadata, and Kubernetes service accounts, pivoting to subsequent systems using discovered credentials.
  3. Activity advances from cloud accounts to data analytics, source-code tokens, and Kubernetes clusters, touching about 10 cloud accounts within roughly 44 hours.
  4. Short-lived branches and workflows are created to output build system secrets into job logs, bypassing standard secret masking via encoding.
  5. Containers that mount the host filesystem as read-only are deployed to collect host credentials, and the creation of a cloud administrator identity was also attempted, though the latter was blocked.

4. Attacker Location and Execution Points

  • External attackers connect from the internet or compromised relays/proxies to public-facing applications, partner/supplier portals, cloud management planes, Microsoft 365, and business APIs.
  • Following initial access, compromised cloud identities, application identities, Windows/Linux servers, and domain controllers can serve as operational vantage points.

5. Visibility for Victims and Administrators

  • Users: When app-only authentication or public application exploitation proceeds without user interaction, clear indicators such as MFA prompts may be absent. Incidents may only be discovered later through unauthorized access to mailboxes, personal data, or business data.
  • Administrators: Activity can be observed through high-frequency sequential requests, dry-runs, test writes, unusual service principal sign-ins, added OAuth credentials, Kudu API executions, web shells, rapid enumeration bursts, staging volume creation, transmissions to external webhooks, resource names starting with i followed by digits, short-lived GitHub branches/workflows, encoded secrets in build logs, and host filesystem mount containers.

6. Success and Failure Conditions

Success Conditions

  • The presence of at least one reachable attack surface, such as an internet-facing application, partner portal, cloud management plane, or leaked credential.
  • The attacker's ability to operate external LLM APIs and ARTEX instances while maintaining network reachability to the target.
  • Post-initial-access accounts or application identities possessing the permissions required for subsequent operations such as enumeration, credential modification, data access, and compute initiation.
  • Reusable credentials left behind in third-party long-lived keys, IaC states, metadata, or service accounts that allow pivoting to subsequent systems.

Failure Conditions and Mitigations

  • Inventory internet-facing assets and management interfaces, restricting access to Kudu/SCM, hypervisors, backup platforms, and cloud management planes to trusted networks.
  • Separate user identities from workload identities and apply the principle of least privilege, enforcing MFA for user authentication and credential rotation as well as application credential change monitoring for workload identities.
  • Revoke keys upon the conclusion of vendor engagements, migrate to short-lived federation, and prevent secrets from exposure in IaC states, repositories, and build logs.
  • Block iterative exploration and lateral movement using WAFs, EDR, rate limiting, and network segmentation, while feeding blocked attempts into follow-up investigations.

7. What Happens Upon Success

  • Cross-environment access to cloud, SaaS, and on-premises identities and data.
  • Persistence established via web shells, RATs, OAuth credentials, SSH keys, and cloud roles.
  • Data collection, staging, and exfiltration from mailboxes, cloud storage, databases, and local systems.
  • The potential to reach privilege acquisition or information theft before initial response measures can take effect due to parallel processing of multiple targets in a short timeframe.
  • In the Barracuda incident, access reached roughly 10 cloud accounts, analytics, source repositories, and Kubernetes. Systems storing customer integration credentials were also reached, though direct evidence of viewing or retrieval was absent, and incomplete logging prevents exclusion.

8. Observable Logs

  • Email: Confirm reads, searches, bulk access, and mass notifications/broadcasts from compromised mailboxes.
  • Proxy / SWG / DNS: Check for traffic to external LLM APIs, unknown LLM gateways/resellers, Tor/SOCKS/residential proxies, Chisel/SoftEther, and rotating webhook services. LLM API traffic alone does not confirm an attack.
  • Endpoint / EDR: Confirm shell execution from web or database processes, JSP/PHP/ASPX web shells, WISP/vshell/AdaptixC2, PowerShell enumeration, Pass-the-Hash/Pass-the-Ticket attempts, and EDR coverage probing.
  • Identity / IdP: Confirm service principal sign-ins, authentication from unknown autonomous systems, hosting providers, or regions, added OAuth application credentials, cloud role changes, password resets, and evidence of Kerberoasting or DCSync.
  • SaaS / Cloud: Confirm Microsoft Graph activity, Entra audits, Azure ARM dry-runs, sub-second blob writes/deletes, Kudu API usage, IAM PassRole, instance profile modifications, unexpected creation of VMs, volumes, Elastic IPs, or SSH keys, alongside i19ghci.yml/i19ci2.yml files, i19ghci-/i19ci2- branches, encoded outputs in build logs, and privileged or host-mount containers.
  • Network: Check for high-frequency sequential requests to login/search endpoints, short bursts transitioning from scans to exploits, C2, and data transfers, as well as the FBI-listed IOCs and Barracuda-listed IPs 180.178.56[.]10, 52.128.238[.]126, 221.146.68[.]107, and 2a09:bac1:7680:3d0::26b:5b. IP addresses represent historical infrastructure at the time of observation and should be evaluated alongside current telemetry.

9. Assessing Attack Success

Confirmed in Public Information

  • Attack Attempts Observed (Success Unconfirmed): The FBI has synthesized mass scans, dry-runs, authentication/privilege checks, and lateral movement attempts across multiple campaigns. Documentation includes Pass-the-Hash attempts blocked by endpoint protection; therefore, attempt counts are not treated as successful compromises.
  • Malware Execution or Authentication Success Confirmed: In the ConnectWise campaign, Microsoft reports that Storm-0940 attempted sign-ins to over 2,250 tenants using ARTEX, with successful authentications and subsequent operations observed in multiple environments. While the FBI summary states this "enabled access to over 2,250 downstream tenants," Microsoft's description refers to sign-in attempt counts, and successful authentication alone does not indicate data access or exfiltration. The figure of over 2,250 is not treated as the number of successful authentications or victims across all targets.
  • Information Theft or Session Compromise Confirmed: In the South Korean financial campaign, personal data exfiltration has been confirmed in related documentation. The figure of approximately 67,000 individuals reflects the scope reported by the respective sources; it does not necessarily mean that every reported case has been independently verified by the affected organizations.
  • Subsequent Compromises Confirmed: Web shells, RATs, added OAuth credentials, cloud role manipulation, and database/mailbox/storage access have been confirmed across multiple incidents. In the ConnectWise campaign, subsequent operations such as Graph reconnaissance, application credential modifications, and privilege changes were also observed across multiple environments. These represent an aggregation of multiple campaigns rather than a single attack chain by one threat actor.
  • Subsequent Compromises Confirmed (Barracuda): Movement from a valid cloud key to roughly 10 cloud accounts, analytics, source repositories, and Kubernetes was confirmed, along with build workflow secret output and host filesystem mount containers. Cloud administrator identity creation was blocked, and viewing or retrieval of customer integration credentials was not confirmed.

Criteria for Internal Evaluation

  • Do not treat scans, dry-runs, or authentication failures as successful compromise; verify token issuance, command execution, credential modification, data reads, staging, and egress on a stage-by-stage basis.
  • Treat high-frequency automation and LLM API usage as auxiliary indicators, supporting them with identity, endpoint, cloud audit, and network telemetry.
  • Do not conclude compromise based solely on a hit against historical IOCs; correlate timestamps, requests, authentication outcomes, processes, and data access.

10. Investigation Playbook

  • Investigation Starting Points: Begin with high-frequency sequential requests, service principal sign-ins from unknown sources, Kudu/Graph/ARM API anomalies, web shells, unusual egress to LLM gateways, and FBI-listed IOCs.
  • Initial Verification: Establish a chronological timeline of target assets, external reachability, authentication methods, execution privileges, initial successful responses, and immediate subsequent identity or cloud modifications.
  • Endpoint/Server Investigation: Preserve web/database child processes, shells/RATs, PowerShell activity, credential access artifacts, new accounts/SSH keys, and staging files/volumes.
  • Identity/Cloud Investigation: Check service principal sign-ins, OAuth credentials, role assignments, Graph/ARM/Kudu/IAM activity, cloud storage reads, unexpected compute instances, vendor keys, IaC states, metadata, Kubernetes service accounts, and GitHub workflows/branches.
  • Subsequent Operations: Trace internal enumeration, LAPS/Kerberos/DCSync, WinRM, mailbox/database/blob access, and egress via webhooks or proxies.
  • Containment: Revoke compromised identities and application credentials, restrict external attack surfaces and management interfaces, and isolate web shells/RATs/tunnels. Rotate keys, tokens, and passwords after preserving evidence.
  • Classification Categories: Differentiate reconnaissance, initial access, successful authentication, command execution, persistence, lateral movement, data access, staging, and exfiltration.

11. Defense and Detection Ideas

  • Single Events: Treat credential additions to high-privilege applications, arbitrary command execution via Kudu, unexpected role assignments, creation of JSP/PHP/ASPX files in web roots, short-lived workflows that output secrets to build logs, and containers that mount the host filesystem as high-priority items.
  • Chronological Correlation: Correlate valid vendor keys -> secret store/IaC state/metadata reads -> source tokens -> short-lived workflows -> Kubernetes secret/host credential access alongside the chain extending from scans/exploits to staging/egress.
  • Threat Hunting: Cross-search public assets, partner trusts, service principals, Kudu/SCM, ARM dry-runs, blob test writes, new 100GB volumes, and unknown webhooks or LLM proxies.
  • Log Limitations: Because ARTEX can operate on external infrastructure, tool names or model names may not remain in the victim environment. The absence of visible LLM API communication does not prove the absence of ARTEX.
  • Priority Mitigations: Prioritize external attack surface reduction, network restriction of management interfaces, workload identity least privilege, pre-collection of API/Graph/cloud audits, credential modification detection, rate limiting, and segmentation.

12. Facts / Inference / Hypothesis

Facts

  • The FBI reported that since at least August 2026, multiple threat actors have abused ARTEX across global financial, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications, and legal sectors.
  • ARTEX is a multi-agent platform implemented in Go that does not embed its own model, instead integrating with external models such as Claude, GPT, DeepSeek, GLM, Grok, and Kimi.
  • Observations compiled by the FBI include cloud credential privilege verification, exploitation of SAML, XXE, SSRF, SQL injection, and SSTI, web shell/RAT deployment, cloud credential additions, Kerberoasting/DCSync, mailbox/storage/database access, and proxy/tunnel utilization.
  • Some intrusions were reported to have progressed from targeting to privileged access in under three hours, though this duration does not apply to all campaigns.
  • The original repository was made private on October 8, 2026, but forks, archives, localized variations, and ARTEX PRO remain available.
  • The figure of over 1,600 publicly accessible ARTEX consoles identified as of early November 2026 is based on researcher internet observations and does not represent confirmed abuse of all instances by the FBI.
  • FBI IOCs include historical infrastructure related to South Korea, ConnectWise, and Storm-3358, and analysts are cautioned to account for cloud and virtual infrastructure reassignment by verifying with current telemetry.
  • In the Barracuda incident, intrusion occurred via cloud access keys left in a third-party AI vendor's GitHub configuration file, credential-chaining to roughly 10 cloud accounts, analytics, source repositories, and Kubernetes within approximately 44 hours.
  • Attackers used short-lived branches/workflows to output build secrets into encoded logs, bypassing standard secret masking.
  • Credential collection via containers mounting the host filesystem read-only was confirmed, but cloud administrator identity creation was blocked.
  • Access to systems storing customer integration credentials was confirmed, but direct evidence of credential viewing or retrieval was absent, and incomplete logging prevents complete exclusion.

Inference

  • The danger of ARTEX lies less in any specific model itself than in its ability to parallelize repetition across multiple assets, identities, and cloud APIs, thereby compressing the defender's decision window.
  • Because detection based solely on tool names, user-agent strings, or model APIs can be evaded, detection strategies should focus on sequences of privilege checks, credential changes, resource creation, and data access.

Hypothesis

No additional hypotheses. Unverified items are documented in "Unknowns and Additional Investigations."

13. MITRE ATT&CK Mapping

ID Technique Confidence Basis
T1595 Active Scanning high Directory fuzzing and active vulnerability scans have been observed.
T1190 Exploit Public-Facing Application high Exploitation of XXE, SSRF, SQL injection, SSTI, and command injection has been synthesized.
T1078 Valid Accounts high Stolen VPN credentials, cloud identities, and application identities have been utilized.
T1505.003 Server Software Component: Web Shell high Deployment via JSP/PHP web shells and MySQL log poisoning has been observed.
T1098.001 Account Manipulation: Additional Cloud Credentials high Credential additions to existing OAuth applications have been observed.
T1090 Proxy high Residential proxies, Tor, SOCKS, and compromised relays have been utilized.
T1530 Data from Cloud Storage high Mass reads and data staging from cloud storage have been observed.
T1550.002 Use Alternate Authentication Material: Pass the Hash high Lateral movement attempts. The FBI table notes blocking by endpoint protection.
T1550.003 Use Alternate Authentication Material: Pass the Ticket high Lateral movement attempts. The FBI table notes blocking by endpoint protection.
T1021.006 Remote Services: Windows Remote Management high Lateral movement attempts. The FBI table notes blocking by endpoint protection.
T1557.001 Adversary-in-the-Middle: LLMNR/NBT-NS Poisoning and SMB Relay medium The FBI TTP table maps this technique to lateral movement activity. This corresponds to the Kerberos relay described in ยง3, but execution of LLMNR/NBT-NS poisoning or SMB relay itself was not individually confirmed.
T1552.001 Unsecured Credentials: Credentials In Files high Credentials were acquired from third-party GitHub configurations, IaC states, and build logs.

14. Unknowns and Additional Investigation Items

  • Among the over 1,600 publicly accessible consoles, the exact count of malicious operations, legitimate security tests, and unused instances.
  • The complete mapping of which campaigns, attackers, and targets correspond to each TTP compiled by the FBI.
  • The boundary between commands generated and executed by ARTEX versus operations performed directly by human operators.
  • The number of successful compromises, data exfiltration scopes, and independently verified ranges per campaign for victims.
  • Whether Barracuda customer integration credentials were actually viewed or retrieved, the complete vector of the third-party vendor compromise, and operator attribution for the Barracuda incident.

15. Impact on SOCs and Organizations

Organizations and SOC teams should inventory not only public-facing applications but also partner/supplier portals, legacy business APIs, Kudu/SCM, backups/hypervisors, and workload identities as a unified attack surface. Correlate rapid iterative reconnaissance, cloud privilege checks, application credential changes, unexpected compute and storage creation, data staging, and outbound traffic across the organization, without relying solely on ARTEX-specific signatures.

16. Summary by Target Audience

  • For SOCs: Validate FBI IOCs as historical indicators and chronologically correlate high-frequency requests, service principal sign-ins, Kudu/Graph/ARM/IAM activity, credential modifications, staging, and egress.
  • For Administrators: Inventory external-facing surfaces and management interfaces, apply least privilege to workload identities, and enable activity logging for application credentials, roles, and cloud resources.
  • For Users: This incident includes vectors that proceed without user interaction. Report unexpected authentications, account modifications, or business service anomalies, and prioritize cloud and application investigations by administrators.

Top comments (0)