ARTEX: Public GitHub Secrets Became Entra and Azure Access
Microsoft observed ARTEX-assisted searches of public GitHub secrets and subsequent abuse of valid Entra application credentials. Across different victims, follow-on activity included cloud discovery, Key Vault and storage access, Microsoft Graph operations and AI-resource probing. One observed agent progressed from repository search to Key Vault exfiltration in five minutes. The report does not establish one identical end-to-end chain for every tenant.
Credential ExposureService Principal TakeoverCredential GovernanceCloud IAMAgentic AI
A usable secret is already an access path: Public code searches can yield an application identity whose existing grants lead directly to cloud resources. One observed path took five minutes; that does not describe every victim.
Separate discovery from impact: Corrected ARM calls, Graph activity, secret reads and AI-resource probes are different operations. A successful sign-in or metadata request alone does not prove every downstream resource was compromised.
Authentication layers differ: Application-only credentials can work without interactive user login. Removing a public file does not revoke the secret, and removing a secret does not necessarily invalidate an issued token.
Infrastructure as Code (IaC) Security Scanning for AI Systems
High
For AI workload configuration and IaC, block exposed secrets before merge or upload by scanning the current tree, full reachable history, rendered plans and build/configuration artifacts. Include explicit reviewed exclusions and fail closed on scanner errors. GitHub push protection and public monitoring complement the gate but do not invalidate previously copied credentials.
Revoke or rotate exposed application credentials and every secret the affected identity could read, using actual issuer/verifier inventories. Search history and copies, but do not mistake removal from GitHub for revocation. Prove old values fail and replacements reach only trusted consumers. If investigation confirms attacker-added client secrets or certificates, remove those exact credential objects separately from rotating the original secret. Preserve object IDs and audit evidence, then re-enumerate the application to prove no targeted object remains usable.
Treat already-issued access tokens and sessions separately from client-secret removal. Use supported revocation or relying-service denial and maintain containment until verified expiry where immediate invalidation is unavailable. Microsoft explicitly warns that issued tokens can outlive credential removal.
AI Agent & Workload Principal and Issuance Disablement
High
Disable the compromised service principal’s fresh-authentication and federation paths at their authoritative Entra objects. Verify exact subject and binding state, including attacker-added federated credentials where present. Test that fresh issuance fails; do not claim this invalidates tokens already held by the attacker.
Keep application secrets in a managed store with narrow resource scope and short issuer-enforced lifetime, and migrate to workload-native authentication where actually supported. Verify that source/configuration exports omit reusable values and that old credentials fail at Entra. A short expiry only narrows the window; it does not stop the observed minutes-scale misuse. For an affected Azure OpenAI resource, use an approved Entra workload identity with resource-scoped OpenAI User permission and disable local key authentication at the provider. Independently verify configuration and wait for data-plane propagation: an old key must receive authorization rejection while the approved workload succeeds. This bounds AI-service access; it does not protect unrelated Graph or Key Vault permissions.
Correlate Entra service-principal issuance/sign-ins with ARM, Key Vault, Storage, Graph and AI-resource events. Look for unapproved workload/resource combinations, cross-resource fan-out and credential changes, including failures followed by success. Validate event coverage before analysis and keep missing Graph or vault records as an evidence gap.
Honey Data, Decoy Artifacts & Canary Tokens for AI
Medium
An authorized defender can place a uniquely registered, non-production canary credential in an intentionally exposed decoy configuration and monitor its use. Keep the decoy isolated from real data and verify the whole event-to-alert path. The Framework’s AWS example is not an Azure credential format; an Azure-oriented deployment needs a reviewed canary identity and its actual sign-in/alert adapter.
Restore application and AI-service permissions only after credential, issuance and resource-access checks succeed. Start with minimum required resources and monitored traffic, and verify replacement identities cannot reintroduce the old broad access. Recovery of a tenant is not established by a new secret alone.
An authorized cloud defender can expose a dedicated decoy AI endpoint through an owned decoy configuration. Return static synthetic responses, with no real model, credentials, customer data or route into production. Register the endpoint and expected probe events outside the decoy, enforce resource and outbound limits, and test that a controlled probe reaches the responder and monitoring team. This can reveal or divert AI-resource reconnaissance; it does not guarantee that an agent selects the decoy or stops using a stolen real credential.
What Defenders Should Do Now
Scan AI deployment configuration, history and generated artifacts, block supported secrets before publication, and use managed, narrowly scoped workload credentials. Treat any already-public value as compromised.
Correlate Entra and resource-side evidence; investigate behavior beyond published IPs. Revoke exposed and attacker-added credentials, stop fresh issuance, and separately contain existing tokens. Check ownership and licensing before relying on workload Conditional Access.
For Azure OpenAI, verify provider-side identity and local-key settings on the actual data plane. Use isolated canaries only with a tested alert path, and restore application permissions in verified stages.
Conclusion
The deployable defenses are at the victim’s repository, identity and resource boundaries. They reduce the attacker’s access regardless of how quickly its AI agents search or retry.