ARTEX AI Attacks: How to Check Microsoft 365, Entra ID and Azure for Stolen App Credentials

当ページのリンクには広告が含まれています。
Anime-style Japanese female cloud security engineer investigating exposed cloud secrets
ノートンストア

Microsoft 365 and Azure administrators should audit application identities—not only employee passwords. On November 5, 2026, Microsoft Threat Intelligence reported that threat actors had used the agentic AI offensive framework ARTEX to search public GitHub repositories for exposed credentials and abuse valid application secrets to access Entra ID, Azure and Microsoft 365 resources.

In one case observed by Microsoft, an AI agent went from searching GitHub for client secrets to exfiltrating data from Azure Key Vault in just five minutes. That is a documented example, not evidence that every attack takes five minutes or that every Microsoft 365 tenant was compromised.

This English guide adapts our original Japanese report using Microsoft’s November 5 research and its November 6 follow-up. Checked November 7, 2026. We cannot determine an individual organization’s compromise without its logs and inventory.

広告・PR
目次

Quick checklist: what should an Entra ID administrator inspect?

AreaImmediate action
Service principal sign-in logsInvestigate successful and failed application sign-ins from October 28 onward, within the available retention window.
Enterprise applicationsIdentify unfamiliar integrations, broad Graph permissions and administrator consent grants.
App registrationsReview client secrets, certificates, federated credentials and unexpected changes.
SharePoint, OneDrive and ExchangeCorrelate application sign-ins with file downloads, mailbox item access and audit events.
GitHub and other code hostsSearch for exposed secrets; revoke and rotate them. Removing a file alone does not invalidate a leaked credential.
Azure Key Vault and StorageCheck access to secrets, storage keys, blob downloads and suspicious cloud resource enumeration.
Conceptual security diagram showing exposed repository secrets, an AI agent and cloud resources
Concept illustration: a publicly exposed application secret can lead to cloud access within the permissions assigned to that application.

What Microsoft actually observed

Microsoft says it began observing large-scale GitHub secret searches using an agentic framework in August 2026. The searches targeted passwords, application client secrets, API keys and AI-service credentials across public code repositories. The agents used discovered information for follow-on automated operations.

A related pattern of successful service principal sign-ins was observed beginning October 28, 2026. Microsoft investigated access involving multiple application IDs and tenants. Some attacks used compromised credentials belonging to a multitenant vendor application, potentially exposing downstream customer tenants—not merely secrets published by the target organization itself.

Observed behavior included Azure subscription and resource discovery, retrieving Azure Key Vault information, accessing Azure Storage, and activity involving SharePoint, OneDrive and mailbox items. Microsoft observed agents altering cloud API requests after errors and retrying them with different versions or methods. These are observations of attacker behavior, not instructions for reproducing an intrusion.

MilestoneReported information
August 2026Large-volume public GitHub repository secret searching observed
October 28, 2026 onwardSuccessful application/service principal sign-ins observed in some affected tenants
November 5, 2026Microsoft published the research on agentic AI credential abuse
November 6, 2026Microsoft separately described Storm-3358’s ARTEX activity targeting the financial sector
ImpactObserved unauthorized cloud resource and data access in some organizations; not evidence of a universal Microsoft 365 breach

Why user MFA and password resets are not enough

A human employee signs in with an interactive account using methods such as passwords, passkeys and multifactor authentication. A service principal represents an application or automated workload. It can authenticate using a client secret, certificate or federated credential and call APIs directly under permissions granted to that identity.

If an attacker steals an application’s valid credential, they may be able to authenticate as that application without triggering a human user’s ordinary MFA prompt. Changing employees’ passwords alone does not revoke a stolen application secret. The real questions are what credentials were exposed, what permissions the application holds, what resources it accessed, and how long its tokens or secrets remain usable.

ControlHuman user sign-inApplication/service principal sign-in
IdentityEmployee or other userApp or automated workload
Typical authenticationPasskey, password, MFAClient secret, certificate, federated credential
PermissionsUser privileges and delegated permissionsApplication permissions and Azure RBAC grants
Immediate threatStolen user credential or sessionStolen app credential granting programmatic resource access
Priority defenseStrong authentication, conditional access, session reviewCredential protection, least privilege, workload monitoring and rotation

Step-by-step: check Microsoft Entra ID service principal sign-ins

  1. Sign in to the Microsoft Entra admin center with an account authorized to view the relevant logs.
  2. Navigate to Entra ID → Monitoring & health → Sign-in logs, then select the Service principal sign-ins category rather than only the interactive-user log.
  3. Review available records from October 28, 2026 onward, filtering for application ID, source IP, success or failure, resource and timestamp.
  4. Look for unusual bursts of sign-ins across multiple applications, especially unsuccessful attempts followed by successful ones.
  5. Correlate suspicious application identities with API permissions, credential additions, Azure activity and Microsoft 365 data-access events.

Microsoft named 199.19.77.84 and 199.19.77.18 among IP addresses associated with the investigated sign-in activity. The second is a Tor exit node and may also be associated with unrelated activity. Other observed infrastructure includes 65.87.223.26. Treat these as investigation leads, not standalone proof of a compromise; attackers can change IP addresses.

Errors AADSTS7000215 (invalid client secret) and AADSTS7000222 (expired client secret) can be useful indicators when they appear in a relevant sequence, especially if failures are followed by success. Neither error code alone demonstrates malicious activity.

Important: log retention differs by Entra license

Microsoft’s documented standard retention for Entra ID audit and sign-in logs is seven days for Entra ID Free and 30 days for P1/P2. As of November 7, a tenant on Free may no longer have October 28 logs available unless they were exported to another platform. Upgrading later does not restore expired records. These Entra logs are separate from Microsoft 365’s Unified Audit Log and its own retention rules. See Microsoft Learn: Entra data retention.

Anime-style security analyst examining application sign-in alerts and cloud audit records
Illustration: correlate identity, API, storage and Microsoft 365 logs. This is not a real Entra admin screenshot.

Review enterprise application permissions and secrets

1. Inspect enterprise applications

In the Entra admin center, open Entra ID → Enterprise applications → All applications. Review which applications are present, who owns them, when they were added, and whether their Microsoft Graph permissions match their business functions. Review granted administrator and user consent as appropriate. Broad permissions such as Files.Read.All, Sites.Read.All or Mail.Read require analysis of whether they are delegated or application permissions and how narrowly they are scoped.

Do not indiscriminately delete legitimate SaaS integrations or backup agents; doing so can interrupt critical services. Investigate and reduce excessive grants through the organization’s change process.

2. Inspect app registrations

Open Entra ID → App registrations → All applications and inspect Certificates & secrets and API permissions for relevant applications. Check for unexpected client secrets, certificates or federated identity credentials. Remember that a third-party multitenant application’s vendor-owned secret may not appear in your tenant’s own app-registration list. Coordinate with affected vendors where necessary.

3. Investigate data access

Correlate Microsoft Purview audit records (where available), Azure Activity Logs, Key Vault diagnostics, Storage diagnostics, Azure DevOps audit events and Entra audit logs. Investigate unusual file downloads from SharePoint or OneDrive, mailbox item access, secret reads and modifications to app permissions. Absence of a record is not proof no access occurred, especially when auditing was not enabled or retention has expired.

What to do if an application credential is exposed

  1. Preserve evidence: export available sign-in and audit events; record application IDs, service principal object IDs, timestamps, IP addresses and affected resources.
  2. Contain access: revoke or replace compromised secrets, certificates and federated credentials; disable a service principal if justified. Existing access tokens may remain valid until expiry.
  3. Rotate downstream secrets: consider all credentials reachable through a compromised Key Vault, accessed documents, Azure DevOps personal access tokens and other connected systems.
  4. Review privileges: narrow Microsoft Graph application permissions, Azure RBAC and vendor/MSP access.
  5. Restore safely: validate necessary integrations after credential rotation and coordinate with vendors and incident responders.

Microsoft recommends blocking the specific address 199.19.77.84/32 via appropriate Conditional Access controls when applicable. Workload identity Conditional Access requires Workload ID Premium and has additional scope constraints; ordinary user-focused Entra P1 Conditional Access is not equivalent. IP blocking alone will not stop an attacker who changes infrastructure.

Prevent future exposed-secret incidents

  • Enable secret scanning and public repository monitoring where available; review GitHub repositories, CI/CD configuration, code history and unintended copies.
  • Never store long-lived client secrets in public code, configuration files, issues or documentation.
  • Where supported, use managed identities and workload identity federation instead of static application secrets.
  • Periodically inventory enterprise applications, administrators’ grants and service principals; remove unnecessary permissions.
  • Export identity and cloud logs to a monitored storage/SIEM destination if your incident response window exceeds built-in retention.
  • Protect backup and recovery resources from application credentials with broad or unnecessary write access.

Separate November 6 report: ARTEX and financial-sector attacks

On November 6, Microsoft separately reported that the financially motivated group it tracks as Storm-3358 had used ARTEX while targeting financial organizations, particularly in Southeast Asia. That report describes vulnerability exploitation, remote-access malware and broader attack operations. It is not safe to treat every incident mentioned in both reports as the same intrusion or to assign all related events to one actor. See Microsoft’s November 6 report.

Frequently asked questions

Does the report mean every Microsoft 365 user was breached?

No. Microsoft described observed attacks involving exposed or stolen application credentials and access to particular customer environments. Whether a specific tenant was affected requires investigation.

Will changing all employee passwords solve this?

Not by itself. Review and rotate exposed application secrets and related credentials. User-account steps are necessary only when those accounts or sessions were affected.

Why can’t I see the October 28 events?

Check the selected log category, viewing permissions, date range and license retention. Data outside the retention window may be unavailable unless previously exported.

Can I rule out an attack if the known IP addresses are absent?

No. Source IPs and user agents change. Behavioral correlation—multiple applications, token usage, cloud-resource access and downloads—is more reliable than a single indicator.

Summary and official references

The important new risk is not that an AI tool can guess every password. It is that AI agents can discover already-exposed application secrets and rapidly use the permissions of trusted workloads. Check service principal sign-ins, remove unnecessary application grants, rotate exposed secrets and examine the data those apps could access.

English adaptation checked November 7, 2026. Admin-console paths are based on vendor documentation and were not revalidated in a live tenant for this translation. Anime illustrations are conceptual and do not depict authentic attack screenshots.

よかったらシェアしてね!
  • URLをコピーしました!
  • URLをコピーしました!

この記事を書いた人

はじめまして、「ぴろりん」です。
このブログでは、IT運用管理・PC修理・情報セキュリティ対応・資格学習など、実務で直面した課題とその解決策を発信しています。
同じような悩みを抱える方のお役に立てれば幸いです。

目次