AI Agents Turn Secrets Leaked on GitHub Into Cloud Break-Ins in Minutes
Microsoft says attackers use the ARTEX agentic AI framework to find app secrets in public GitHub repos, then sign in as service principals to raid Key Vault, Storage, SharePoint and Azure AI resources.
Attackers are pointing an AI agent framework at GitHub to dig up leaked app secrets, then letting the agents log in and loot the cloud on their own. Microsoft Threat Intelligence says the tool is ARTEX, the offensive framework it previously tied to Storm-0940’s ConnectWise supply-chain attack. In one case an agent went from searching GitHub for a client secret to pulling data out of Azure Key Vault in five minutes. Microsoft doesn’t name the actors behind this wave.
For defenders
Finding the keys: since August 2026 the framework has fired large volumes of GitHub searches combining a target company’s name with secret, cloud and AI-provider keywords (T1593.003 (Search Open Websites/Domains: Code Repositories), T1552.001 (Unsecured Credentials: Credentials In Files)). It wants app client secrets and API keys for AI services in particular.
Using them: from Oct. 28, agents signed in as service principals for dozens of apps across multiple tenants, mostly from 199.19.77[.]84 (T1078.004 (Valid Accounts: Cloud Accounts)). One stolen secret belonged to a vendor’s multitenant app, which opened its customers’ tenants too. The agents went after anything that returns credentials: storage account keys, App Service settings and publishing credentials, connection strings, AKS credentials and Key Vault secrets (T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores)). They downloaded SharePoint and OneDrive files, read mailbox items, ran Graph searches, searched Azure DevOps for more secrets, and used VM Run Command to execute scripts (T1651 (Cloud Administration Command)).
The agent’s fingerprint: when an Azure Resource Manager call failed on a bad API version, it retried 28 to 80 seconds later with only the api-version changed and got HTTP 200. A second address, 199.19.77[.]18, is a Tor exit node, so Microsoft warns it also carries unrelated traffic. 65.87.223[.]26 sent shell commands inside its user agent and tried to grab App Service tokens; Todyl reported the same IPs. 157.22.126[.]128 scouted Azure AI workloads.
Check your logs: sign-in errors AADSTS7000215 (invalid secret) or AADSTS7000222 (expired secret) mean someone is trying a credential for your app; a failure followed by a success is urgent. Microsoft’s hunting queries are in the Hunt Lead report below.
199.19.77[.]84 (service principal sign-ins, ARM and Graph activity)
65.87.223[.]26 (exploit attempts in user agent, App Service token requests)
157.22.126[.]128 (reconnaissance of Azure AI workloads)What it means for you
This one is about businesses and the developers who build their apps. If a password or key for a company app ever ended up in code on GitHub, assume a robot has already found it; deleting the file doesn’t help, because old copies stay findable until the key itself is changed. For everyday users there’s nothing to do.
Do this now
- Turn on secret scanning and push protection for your GitHub and Azure DevOps repositories.
- Rotate every secret or key that has ever appeared in a repository, public or private.
- Move apps from client secrets to managed identities or certificates where you can.
- Review Entra service principal sign-ins for the listed IPs and for AADSTS7000215/7000222 errors.
Source article: Threat actors use agentic AI tooling to find and abuse exposed secrets
Publisher: Microsoft Threat Intelligence (Microsoft Security Blog)
Published: Nov. 5, 2026
URL: https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/
Executive Summary
Microsoft Threat Intelligence reports that unnamed threat actors have used ARTEX, the agentic AI offensive framework Microsoft previously documented in Storm-0940’s ConnectWise supply-chain compromise, to search public GitHub repositories at scale since August 2026 for exposed secrets: user credentials, OAuth application client secrets and API keys for cloud and AI inference services. The agent’s queries pair a target organization’s name with secret, cloud-service or AI-provider keywords, then the agents analyze what they find and act on it without a human at the keyboard. In the follow-on wave Microsoft tracked from Oct. 28, 2026, agents used leaked or stolen application credentials to sign in as service principals for dozens of applications across multiple Entra ID tenants, chiefly from 199.19.77[.]84 and also from the Tor exit 199.19.77[.]18, including a vendor’s multitenant application whose compromise reached downstream customer tenants; they enumerated subscriptions, resource groups, permissions, AI model deployments and configurations; read storage account keys, App Service settings and publishing credentials, connection strings, AI and ML datastore connections, AKS user credentials and Key Vault secrets; downloaded SharePoint and OneDrive files, accessed mailbox items and ran Microsoft Graph searches; searched Azure DevOps for further secrets; ran scripts on Azure VMs with Run Command; and tried to validate and reuse compromised Azure AI resources. A related address, 65.87.223[.]26, embedded shell commands in its user agent and tried to obtain App Service tokens, and 157.22.126[.]128 performed reconnaissance of Azure AI workloads; Todyl reported a similar campaign from the same IPs. The objective is credential, data and AI-resource theft at machine speed, with persistence through unrotated secrets. Hunt leads, false positive expectations, Microsoft’s published hunting queries and indicators are detailed below.
ATT&CK chain: Reconnaissance → Credential Access → Initial Access → Discovery → Collection → Execution → Exfiltration
Hunt Lead Grid
| Hunt Lead Level | Hunt Lead Name | ATT&CK | Risk of False Positives |
|---|---|---|---|
| 1 (Junior) | Campaign IP Sweep Across Entra, Graph and Azure Logs | T1078.004 (Valid Accounts: Cloud Accounts) | Low for 199.19.77[.]84, 65.87.223[.]26 and 157.22.126[.]128, which Microsoft ties to this activity; 199.19.77[.]18 is a Tor exit carrying unrelated traffic, so a hit on it alone needs corroboration before escalation. |
| 1 (Junior) | Leaked App Secret in Use (AADSTS7000215 and 7000222) | T1552.001 (Unsecured Credentials: Credentials In Files), T1593.003 (Search Open Websites/Domains: Code Repositories) | Medium. A developer with a stale secret in a config file also produces these errors; anchor on failures from IPs outside your egress ranges and on any failure followed by a success from the same new source. |
| 2 (Pro) | ARM API-Version Retry Pattern from a Service Principal | T1580 (Cloud Infrastructure Discovery), T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores) | Medium. SDK upgrades and hand-written scripts sometimes retry with a new api-version; the lead anchors on a batch of HTTP 400s followed within about 90 seconds by the same URIs with only api-version changed and HTTP 200, from a service principal against secret-bearing resources. |
| 3 (Elite) | One Source, Many Service Principals, and Cross-Resource Token Burst | T1078.004 (Valid Accounts: Cloud Accounts), T1651 (Cloud Administration Command) | Medium. Automation platforms and MSP tooling legitimately sign in as many apps from one IP; the lead needs an allowlist of those sources and anchors on mixed success across three or more service principals or one app reaching ARM, Key Vault, Storage and Graph in a short window. |
Campaign IP Sweep Across Entra, Graph and Azure Logs
From Oct. 28, 2026, Microsoft saw successful service principal sign-ins from 199.19.77[.]84 for multiple application IDs across customers, followed by Azure Resource Manager enumeration, Key Vault and Storage reads, SharePoint and OneDrive downloads, mailbox access and Graph searches from the same address; 199.19.77[.]18, a Tor exit, carried some of the sign-ins. 65.87.223[.]26 showed exploitation attempts with shell commands embedded in the user agent and attempts to obtain App Service authentication tokens, and 157.22.126[.]128 performed reconnaissance against Azure AI workloads. The detection logic is a sweep of every identity and cloud log you hold for those addresses, successful and failed, using Microsoft’s published Defender XDR advanced hunting queries below: Entra service principal sign-ins, Graph API audit events, Cloud App events, cloud control-plane audit events and storage operations.
Run the queries in Defender XDR advanced hunting, or translate the tables to their Sentinel equivalents (AADServicePrincipalSignInLogs, MicrosoftGraphActivityLogs, AzureActivity, StorageBlobLogs, KeyVault diagnostics) and add Azure DevOps and Microsoft 365 unified audit logs, which Microsoft also asks you to review. Extend the start date as far back as retention allows; Microsoft’s queries begin at Oct. 28, but the GitHub harvesting began in August. Note that some published queries list only 199.19.77[.]84; add 199.19.77[.]18, 65.87.223[.]26 and 157.22.126[.]128 to your own copies. Block 199.19.77[.]84/32 as a named location in Conditional Access for workload identities and in a user-scoped policy, as Microsoft recommends, and block the exact IPs rather than the ASN. Treat a successful sign-in from these addresses as a compromised application: remove all its secrets, certificates and federated credentials, revoke related sessions, PATs and SSH keys, and rotate everything in any Key Vault it could read. Issued tokens stay valid until they expire, so watch for continued activity after rotation.
let startTime = datetime(2026-10-28T00:00:00Z);
let endTime = now();
EntraIdSpnSignInEvents
| where Timestamp between (startTime .. endTime)
| where IPAddress in ("199.19.77.84")
| project Timestamp, TenantId, Application, ApplicationId, ServicePrincipalName, ServicePrincipalId, ResourceDisplayName, ResourceId, ResourceTenantId, IPAddress, ErrorCode, UserAgent, CorrelationId, RequestId
| order by Timestamp asc
let startTime = datetime(2026-10-28T00:00:00Z);
let endTime = now();
GraphAPIAuditEvents
| where Timestamp between (startTime .. endTime)
| where IpAddress in ("199.19.77.84")
| project Timestamp, TenantId, ApplicationId, ServicePrincipalId, AccountObjectId, IpAddress,
RequestMethod, RequestUri, ResponseStatusCode, ResponseSize, Scopes, EntityType,
TargetWorkload, UniqueTokenIdentifier, ClientRequestId, OperationId, RequestId
| order by Timestamp asc
let startTime = datetime(2026-10-28T00:00:00Z);
let endTime = now();
CloudAppEvents
| where Timestamp between (startTime .. endTime)
| where IPAddress in ("199.19.77.84")
| project Timestamp, Application, CloudAppId = ApplicationId, ActionType, ActivityType,
AccountType, AccountDisplayName, AccountId, AccountObjectId, IPAddress, UserAgent,
ObjectName, ObjectType, ObjectId, ActivityObjects, ReportId, RawEventData
| order by Timestamp asc
CloudAuditEvents
| where IPAddress has_any ('199.19.77.84','65.87.223.26')
CloudStorageAggregatedEvents
| where IPAddress in ('199.19.77.84','65.87.223.26')
| where SuccessfulOperationsCount > 0
(Advanced hunting queries reproduced as published by Microsoft Threat Intelligence.)
Hunt Lead Level: 1 (Junior Analyst)
ATT&CK: Initial Access → T1078.004 (Valid Accounts: Cloud Accounts) (our mapping)
Sources: “Threat actors use agentic AI tooling to find and abuse exposed secrets”, Microsoft Threat Intelligence, Nov. 5, 2026, https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/ (source article)
Leaked App Secret in Use (AADSTS7000215 and 7000222)
The whole chain starts with a secret committed to a public repository: developers put client secrets, storage keys and connection strings in configuration files and scripts to get an OAuth application working quickly, and ARTEX’s GitHub queries combine your organization’s name with words like secret, key, Azure or the name of an AI provider to find them. Microsoft’s point is that a published secret stays usable until it’s revoked or rotated, and deleting the file or the commit does not remediate it. Microsoft gives a cheap tell that someone outside is holding one of your app’s credentials: Entra sign-in errors AADSTS7000215 (invalid client secret) and AADSTS7000222 (expired client secret) on service principal sign-ins, which is what an agent produces when it tries an old or rotated secret; a failure followed by a success from the same source is the urgent case. The detection logic has two parts: an inventory check, running GitHub secret scanning with public monitoring and push protection (or GitHub Advanced Security for Azure DevOps) to list every secret of yours that has ever been exposed, and a log check for those two error codes on service principal sign-ins from IPs outside your own egress ranges and CI providers.
Use the Entra service principal and non-interactive sign-in logs (Sentinel AADServicePrincipalSignInLogs, or EntraIdSpnSignInEvents in Defender XDR) with ResultType 7000215 or 7000222, grouped by application ID and source IP, and GitHub secret scanning alerts for owned repositories plus public monitoring for leaks in other people’s repositories. Search as far back as retention allows. Scope first to multitenant apps you publish, since a leak there exposes your customers, then to apps with Key Vault, Storage or Graph permissions. Expect benign hits from developers and pipelines using a stale secret; allowlist by known source IP and workload, not by app. Hiding the repository is the wrong fix; rotating the secret, then moving the app to a managed identity or certificate credential, is the right one. Any secret confirmed exposed should be treated as compromised and its app’s activity reviewed with the other leads.
Hunt Lead Level: 1 (Junior Analyst)
ATT&CK: Credential Access → T1552.001 (Unsecured Credentials: Credentials In Files); Reconnaissance → T1593.003 (Search Open Websites/Domains: Code Repositories) (our mapping)
Sources: “Threat actors use agentic AI tooling to find and abuse exposed secrets”, Microsoft Threat Intelligence, Nov. 5, 2026, https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/ (source article)
ARM API-Version Retry Pattern from a Service Principal
Microsoft caught the agent troubleshooting itself in a victim’s Azure Resource Manager from 199.19.77[.]84. A batch of requests to list secret metadata across several Key Vaults used a Key Vault data-plane API version against the ARM control plane and returned HTTP 400; about 28 seconds after the first failure, the same vault paths were requested again with api-version 2023-07-01 and returned HTTP 200. A GET for an Azure SQL server’s firewall rules with an outdated api-version failed and was repeated about 60 seconds later with a supported version, the URL differing only in that parameter; an Application Insights query POST did the same about 80 seconds after its failure. Human operators and fixed scripts rarely correct themselves this way at this speed across several resource types. The detection logic is to group ARM operations by caller (service principal or app ID) and normalized request path with the api-version parameter removed, then flag a path that returned HTTP 400 and, within about two minutes, returned HTTP 200 with a different api-version, especially for Key Vault, SQL firewall, Storage key or App Service configuration operations.
Use Azure Activity logs where they carry HTTP status and request details, Defender for Resource Manager alerts, and the CloudAuditEvents table in Defender XDR; ARM request telemetry with full URIs gives the cleanest result. Sweep from August 2026 forward. Scope to service principals, since the activity ran under application identities, and weight it by what the resource returns (secrets, keys, firewall rules). Expect some noise when teams upgrade SDKs or hand-test APIs; suppress known deployment identities by app ID and source IP, never by resource. Blocking specific api-versions is the wrong control. Treat a hit from an unfamiliar source as an application compromise and pivot with the IP sweep and the multi-service-principal lead. The pattern generalizes to any LLM-driven tool working through cloud APIs by trial and error.
Hunt Lead Level: 2 (Pro)
ATT&CK: Discovery → T1580 (Cloud Infrastructure Discovery); Credential Access → T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores) (our mapping)
Sources: “Threat actors use agentic AI tooling to find and abuse exposed secrets”, Microsoft Threat Intelligence, Nov. 5, 2026, https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/ (source article)
One Source, Many Service Principals, and Cross-Resource Token Burst
Because the agents worked through a pile of harvested secrets, the same address signed in as dozens of service principals across tenants, some successfully and some not, and each compromised application then requested tokens for several resources in quick succession: ARM to enumerate subscriptions, resource groups, role assignments and AI model deployments; Key Vault and Storage to pull secrets and keys; Graph for SharePoint, OneDrive and mailbox content and searches; and in some cases VM Run Command to execute scripts inside Azure virtual machines. Microsoft’s own behavioral guidance is the anchor: one address authenticating to three or more service principals with mixed success, or one app requesting tokens across Azure Resource Manager, Key Vault, Storage and Graph in a short window. The detection logic is two stateful counts over service principal sign-ins, distinct service principals per source IP per hour with at least one failure and one success, and distinct resource audiences per app ID within a short window, joined to follow-on high-impact operations (Key Vault secret gets, storage listKeys, App Service publishing credential reads, Run Command invocations, or credentials added to the app).
Use Entra service principal sign-in logs with ResourceDisplayName or ResourceId, Azure Activity logs for Run Command and listKeys operations, Key Vault diagnostics, Graph activity logs and Entra audit logs for credential additions; Microsoft also flags “Suspicious service principal sign-in following credential addition” as an alert worth correlating. Sweep back as far as retention allows. Scope to service principals with access to secrets, storage or VMs, and to multitenant apps. Maintain an allowlist of automation platforms, MSP tools and CI egress IPs that legitimately sign in as many apps, keyed by IP and app ID, and review it regularly, because that allowlist is what makes this lead quiet. Disabling Run Command or Graph for everyone is the wrong reaction; least-privilege role assignments for each app and Conditional Access for workload identities limit what a stolen secret can do. A correlated hit should go to incident response with a full credential rotation for the app and every secret it could read; MSPs should also review Partner Center, GDAP relationships and customer tenants, per Microsoft.
Hunt Lead Level: 3 (Elite)
ATT&CK: Initial Access → T1078.004 (Valid Accounts: Cloud Accounts); Execution → T1651 (Cloud Administration Command) (our mapping)
Sources: “Threat actors use agentic AI tooling to find and abuse exposed secrets”, Microsoft Threat Intelligence, Nov. 5, 2026, https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/ (source article)
Hash Appendix
The source article publishes no file hashes; this activity used stolen credentials and cloud APIs, not malware files.
References Appendix
- “Threat actors use agentic AI tooling to find and abuse exposed secrets”, Microsoft Threat Intelligence, Nov. 5, 2026, https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/ (source article; activity description, indicators, mitigations and advanced hunting queries)
- Todyl blog post on stolen app credentials used to sign in to Entra ID and Azure DevOps, Todyl, https://www.todyl.com/blog/stolen-app-credentials-real-access-an-actor-is-signing-into-entra-id-and-azure-devops-as-your-workloads (cited by the source article as reporting the same IPs; not used for additional detail)
- MITRE ATT&CK Enterprise technique pages linked above, MITRE, https://attack.mitre.org/