Threat feed
Hot threat

ARTEX Key Eater cloud credential sweep

ARTEX agents signed in with stolen Entra app credentials to sweep MSP client tenants

Posted · Spectrum Delta Research Team

Summary

ARTEX, the AI penetration-testing framework used by criminal and state actors, powered an unattributed campaign that signed in to MSPs' Microsoft 365 and Azure environments as their own applications using stolen app credentials, reaching 27 organizations through more than 33 compromised app registrations (Microsoft, A/1; Todyl, B/1). Microsoft reported on November 5 that ARTEX agents search public GitHub for exposed secrets, then read Key Vault secrets, storage keys and App Service credentials and use VM Run Command. Todyl traced the same operator since November 1 across roughly 18 organizations, searching about 25 tenants, downloading 81 files and reading about 2,279 mailbox items, and VulnCheck caught the same infrastructure exploiting Langflow CVE-2025-3248. AWS assesses with moderate confidence that Chinese-nexus groups ran ARTEX campaigns from October 3 to November 5 harvesting keys from exposed LiteLLM, One API, Dify and Langflow gateways, and Darktrace reported an October 20 ARTEX-consistent Azure sweep that listed keys and changed firewall rules. An FBI FLASH alert followed on November 13; no Microsoft vulnerability is involved and the operator remains unattributed.

Breakdown

Targets

  • Industry: Managed service providers
  • Industry: Government
  • Industry: Energy
  • Industry: Food and agriculture
  • Industry: Transportation
  • Industry: Construction and engineering
  • Industry: Legal
  • Industry: Education
  • Industry: Nonprofit

Technical exposure

  • Entra app or service principal secret exposed in code and still valid
  • MSP multi-tenant app or delegated-admin identity with tenant access
  • Key Vault, storage keys and App Service credentials readable by apps
  • Azure VM Run Command permitted for application identities
  • Internet-reachable Langflow instance holding AI provider keys
  • Exposed AI gateways with upstream provider keys or default admin

Adversary

Key Eater operator (unattributed, ARTEX agentic framework) · Microsoft, Todyl, VulnCheck, AWS and Darktrace

Procedures (12)

  1. Service principal signs in with a stored secret from a hosting address

    Logs

    • Entra ID service principal sign-in logs (AADServicePrincipalSignInLogs) · ClientCredentialType sign-ins by service principals from AS203662 hosting addresses in 199.19.77.0/24, often an invalid-secret failure preceding a success for the same app; user agents task177-f2267/1.0 and sait-tenant-deep/37346
  2. Partner multi-tenant app enters customer tenants from a hosting address

    • T1199Trusted Relationship

    Logs

    • Entra ID sign-in logs (SigninLogs) · app-only sign-ins where the application is owned by another tenant (MSP partner) or a delegated-admin GDAP identity, from a hosting-provider ASN; one app swept 15 client tenants in about 10-second loops
  3. User refresh tokens replayed by a scripted client

    • T1550.001Use Alternate Authentication Material: Application Access Token

    Logs

    • Entra ID non-interactive user sign-in logs (AADNonInteractiveUserSignInLogs) · refresh-token grants for several users from one hosting address with Python-urllib or curl user agents within minutes, mostly blocked by Conditional Access; one Azure DevOps user session was also replayed
  4. Application identity reads secrets from a cloud key vault

    • T1555.006Credentials from Password Stores: Cloud Secrets Management Stores

    Logs

    • Azure Key Vault diagnostic logs (AuditEvent) · SecretGet and SecretList operations by an application principal from a new address
    • Entra ID service principal sign-in logs · a Key Vault token issued to the app right after an ARM token; Microsoft saw one case go from GitHub secret search to Key Vault exfiltration in five minutes
  5. Application identity pulls resource keys through management APIs

    • T1552Unsecured Credentials

    Logs

    • Azure Activity log · storage account listKeys, web app config list or publishxml, connection strings, AI/ML connections and AKS listClusterUserCredentials called by a service principal from an outside address
  6. App-only identity bulk-downloads SharePoint and OneDrive files

    • T1213.002Data from Information Repositories: Sharepoint

    Logs

    • Microsoft 365 unified audit log (SharePoint / OneDrive) · FileDownloaded operations by app@sharepoint at about one per second, favouring names like pass, key and access, often right after a wildcard Graph search
  7. Application identity reads mailbox items through the Graph API

    • T1114.002Email Collection: Remote Email Collection

    Logs

    • Microsoft 365 unified audit log (Exchange) · MailItemsAccessed operations by an app-only identity from a hosting address (about 2,279 items per Todyl)
  8. Application identity downloads blobs from cloud storage

    • T1530Data from Cloud Storage

    Logs

    • Azure Storage blob logs (StorageBlobLogs) · GetBlob operations by an application identity or account key from a new hosting address
  9. Application identity runs scripts on a cloud VM through Run Command

    • T1651Cloud Administration Command

    Logs

    • Azure Activity log · VM run command operations invoked by a service principal from an outside address
  10. Compromised AI app process queries the cloud metadata service

    • T1552.005Unsecured Credentials: Cloud Instance Metadata API

    Logs

    • EDR network connection events · a Python web-application process connecting to 169.254.169.254 for metadata and user-data and querying the Docker socket after Langflow CVE-2025-3248 exploitation (|RAISE| exception prefix with def _lf_probe() in validate/code POSTs)
    • Linux audit (connect syscall) · auditd connect syscalls to 169.254.169.254 from the application interpreter
  11. Outside client harvests provider keys from exposed AI gateways

    • T1552.001Unsecured Credentials: Credentials In Files

    Logs

    • Web server / reverse proxy access logs in front of the AI gateway · 200 responses for environment files, source maps and debug routes to an unfamiliar client, plus management-API reads of provider configuration endpoints (LiteLLM, One API default admin credentials, anonymous management endpoints, One API isCliRequest header bypass)
    • AWS CloudTrail management events · with stolen keys, identity validation through token and identity services followed by storage and permission enumeration
  12. Cloud identity changes firewall or security rules through the management API

    • T1562.007

    Logs

    • Azure Activity log · network security group and firewall rule writes by a compromised application or user identity from an unrecognized address (Darktrace October 20 sweep)

Indicators of compromise

IPs (7)

  • 199.19.77.84operator VPS: app sign-ins, file downloads and token replays (Todyl, Microsoft)
  • 199.19.77.18sign-in source; Tor exit node, expect unrelated noise (Microsoft)
  • 65.87.223.26App Service token attempts with shell-injection user agent (Microsoft)
  • 157.22.126.128Azure AI workload reconnaissance (Microsoft)
  • 199.19.77.0/24AS203662 range holding operator and spray addresses (Todyl)
  • 104.28.225.223ARTEX AI-gateway key theft (AWS); hunt-only
  • 43.98.196.73Alibaba Cloud address opening the October 20 ARTEX-consistent Azure sweep (Darktrace); hunt-only

CVEs (1)

  • CVE-2025-3248Langflow RCE exploited by the same operator to steal AI provider keys (VulnCheck); KEV listed

Sources

Questions, detections or sightings of ARTEX Key Eater cloud credential sweep? Bring them to the community.