
AI Is Hunting for Your Exposed Secrets Before Your Security Team Finds Them
Exposed API keys and application credentials have always been dangerous. Agentic AI is changing the speed of the threat. Microsoft Threat Intelligence says attackers are now using AI-powered offensive frameworks to search public GitHub repositories for credentials, analyze what they unlock, authenticate to cloud environments, and continue the attack chain with far less hands-on-keyboard work. In one Microsoft-observed case, an AI agent progressed from finding a client secret in GitHub to exfiltrating data from Azure Key Vault in five minutes.
By CompuAce Team
Published November 9, 2026
Reviewed against Microsoft, GitHub, Microsoft Learn, and OWASP guidance as of November 9, 2026
The Short Version
On November 5, 2026, Microsoft disclosed that threat actors were using the agentic offensive security framework ARTEX to submit high volumes of GitHub searches for exposed secrets. The searches included organization names, cloud-related keywords, application credentials, and API keys for AI services.
Microsoft observed follow-on activity where leaked or stolen application credentials were used to authenticate as Microsoft Entra service principals and access connected Azure and Microsoft 365 resources. Attackers enumerated cloud environments, queried Azure Key Vault, accessed Storage Accounts, searched SharePoint and OneDrive, and interacted with Microsoft Graph and Azure resource-management APIs.
The most important operational lesson is simple: if a real secret was exposed publicly, removing the line from GitHub is not remediation. The credential must be revoked or rotated and its historical use investigated.
A cloud credential posted publicly should be treated as compromised, even if it was visible only briefly. Automated systems can find, classify, validate, and abuse exposed secrets at machine speed.
What Microsoft Observed
Microsoft says this activity began at least as early as August 2026, when threat actors started using an agentic AI framework to issue large volumes of GitHub searches designed to discover authentication material. Those searches often combined a target organization name with terms related to secrets, cloud providers, or AI services.
The important shift is not that searching public code for secrets is new. As we covered in AI Agents Are Your Newest Insiders, autonomous tools change who, or what, can act inside your environment. Attackers have done that for years. The shift is that an agentic system can perform multiple steps of the workflow itself: search repositories, analyze results, identify which provider a credential belongs to, test access, adapt when an API call fails, and proceed to follow-on cloud actions.
From a GitHub Secret to Azure Key Vault in Five Minutes
One of the most striking details in Microsoft's report is the speed of an observed attack. In one case, an AI agent moved from finding a client secret in GitHub to exfiltrating data from Azure Key Vault in approximately five minutes.
That does not mean every exposed secret will be exploited in five minutes. It does mean traditional assumptions about response windows are becoming less comfortable. A credential can move from accidental disclosure to active cloud access faster than a human security team may notice the original commit.
What Counts as an Exposed Secret?
In security, a “secret” is authentication material that can grant access to an application, service, workload, or data source. Common examples include:
- OAuth application client secrets
- Microsoft Entra service principal credentials
- Cloud API keys
- Azure Storage account keys
- Database usernames and passwords
- Connection strings
- GitHub personal access tokens
- SSH private keys
- CI/CD credentials
- AI provider API keys
- Application publishing credentials
- Secrets embedded in
.env, configuration, script, or deployment files
A secret can be exposed even when developers never intended to publish it. One committed configuration file, old branch, test script, issue attachment, copied example, or forgotten repository can be enough.
How the Attack Chain Can Work
A simplified version of the Microsoft-observed pattern looks like this:
- Discovery: An agent searches public repositories for likely secrets associated with a target.
- Classification: The framework identifies whether a credential belongs to Azure, Microsoft Entra, an AI service, storage, or another platform.
- Validation: The attacker tests whether the credential is still active.
- Authentication: A valid application credential can be used to sign in as a service principal or OAuth application.
- Cloud discovery: The agent enumerates subscriptions, resource groups, permissions, storage, Key Vaults, AI deployments, and other services.
- Credential access: It searches for additional keys, connection strings, tokens, publishing credentials, or secrets stored in accessible resources.
- Collection or exfiltration: Data may be downloaded from Storage, Key Vault, SharePoint, OneDrive, or other authorized resources.
- Expansion: Additional credentials can enable persistence, lateral movement, or access to more applications.
Why Service Principal Credentials Are So Valuable
A Microsoft Entra service principal is the identity an application uses to authenticate and access resources programmatically. Unlike a human employee, an application identity may operate continuously, without interactive sign-in or MFA prompts in the way people typically experience them.
The risk therefore depends heavily on permissions. If the service principal can read Key Vault secrets, access Storage, query Microsoft Graph, run commands on Azure VMs, or read company files, a stolen credential may grant the attacker the same authorized access until that credential is disabled, revoked, rotated, or otherwise blocked.
Workload identities should have only the permissions their application needs. A leaked credential with narrowly scoped access has a smaller blast radius than a broadly privileged one.
Deleting the Secret From GitHub Is Not Enough
This is one of the most important points in both Microsoft's and GitHub's guidance. Removing a secret from the visible source file does not make the credential invalid.
The exposed value may already exist in:
- Git commit history
- Forked repositories
- Developer clones
- Search indexes
- Build logs
- CI/CD artifacts
- Caches
- Archives and backups
- Automated attacker datasets
The correct response is to revoke or rotate the credential first, then clean the code and history as appropriate, and investigate whether the credential was used while exposed.
Why .gitignore Is Not a Secrets Management Strategy
.gitignore is useful for preventing selected files from being added to Git, but it is not a complete security control. It does not rotate credentials, detect secrets already committed, prevent a developer from pasting a key into another file, or protect secrets that appear in logs, issues, documentation, or build output.
Strong secrets management requires multiple controls: secret detection, push protection, centralized secret storage, least privilege, rotation, monitoring, and architectures that reduce the need for long-lived secrets in the first place.
How GitHub Secret Scanning and Push Protection Help
GitHub Secret Protection includes secret scanning and push protection capabilities designed to identify supported credentials and prevent many accidental leaks.
Secret scanning
GitHub scans public repositories for supported secret patterns. Depending on plan and repository type, organizations can enable secret scanning for private and internal repositories as well. GitHub also supports enterprise monitoring for credentials exposed in public repositories beyond repositories the enterprise directly owns.
Push protection
Push protection can stop a supported secret before it reaches the repository by blocking a push or commit. GitHub supports this protection across command-line pushes, GitHub's web interface, file uploads, relevant REST API operations, and other supported paths.
No scanner catches every possible credential format, so secret scanning should be treated as one layer of defense, not permission to store long-lived credentials in source code.
The Better Fix: Reduce the Number of Secrets You Need
Microsoft specifically recommends moving supported Azure workloads away from secret-based application authentication toward managed identities where possible.
Managed identities allow supported Azure resources to obtain Microsoft Entra access tokens without developers creating, copying, storing, or rotating a client secret in application code. If there is no static secret in the repository, that particular secret cannot be leaked from the repository.
Managed identities are not a universal replacement for every credential, and they still require carefully scoped authorization. But they illustrate an important security design principle: the safest long-lived secret is often the one your application does not need to possess.
Why Agentic AI Changes the Economics of Secret Hunting
Traditional secret hunting required substantial manual effort: find repositories, identify likely credentials, understand the provider, test access, interpret errors, adjust requests, and decide what to do next.
Microsoft's investigation shows that agentic frameworks can automate more of this chain. The observed agents adapted API requests after failures, changed API versions, retried actions, enumerated resources, and continued operating at machine speed.
The security implication is not that AI invented credential theft. It is that AI can reduce the cost, skill, and time required to turn an exposed credential into useful access.
What to Do in the First 30 Minutes After Finding an Exposed Cloud Secret
- Assume the secret is compromised. Do not wait for proof of misuse before containing it.
- Revoke or rotate it immediately. Removing the code alone is not enough.
- Identify the owning identity. Determine whether it belongs to a service principal, user, CI/CD process, database, storage account, or external provider.
- Map its permissions. Identify every resource the credential could access.
- Review sign-in logs. Look for unusual service principal, workload identity, or non-interactive authentication.
- Review cloud audit logs. Check Key Vault, Storage, Azure Resource Manager, Microsoft Graph, Microsoft 365, and other relevant services.
- Look for secondary secrets. If the identity could read a Key Vault, storage configuration, documents, or deployment settings, assume additional credentials may have been exposed.
- Rotate downstream credentials when justified.
- Reduce excessive permissions before restoring access.
- Document the incident and prevent recurrence.
What Security Teams Should Review
Microsoft recommends looking beyond a single indicator such as an IP address because agentic attack tooling can change infrastructure and user-agent characteristics quickly. Useful investigation areas include:
- Microsoft Entra service principal and non-interactive sign-in logs
- Repeated authentication attempts across multiple applications
- Failed credential attempts followed by successful sign-ins
- Azure Activity logs
- Azure Key Vault diagnostics
- Azure Storage access
- Microsoft 365 unified audit logs
- Microsoft Graph activity
- Azure DevOps audit logs
- Unexpected PAT creation or service connections
- Cloud-resource enumeration followed by secret reads or downloads
Build a Real Secrets Lifecycle
OWASP recommends treating secrets as assets with a lifecycle: creation, rotation, revocation, and expiration. For businesses, that means every important credential should have clear ownership and a defined process for replacement.
A mature secrets program should answer:
- Who owns this secret?
- Which application uses it?
- What resources can it access?
- Where is it stored?
- When was it created?
- When does it expire?
- How is it rotated?
- How quickly can it be revoked?
- Which systems break if it changes?
- How would we detect unauthorized use?
The 15-Point Exposed Secrets & Cloud Identity Checklist
- Scan public repositories. Search for real credentials and historical exposures.
- Scan private code too. Internal repositories can later become public, copied, or compromised.
- Enable secret scanning. Use GitHub or equivalent source-code security tooling.
- Enable push protection. Stop supported secrets before they reach repositories.
- Treat exposed secrets as compromised.
- Revoke before cleaning history. Make the credential useless first.
- Inventory service principals and OAuth applications.
- Review workload identity permissions. Apply least privilege.
- Replace static secrets with managed identities or federation where supported.
- Use a dedicated secrets manager. Avoid hardcoded production credentials.
- Prefer short-lived credentials. Reduce the useful lifetime of stolen authentication material.
- Monitor Key Vault and secret-store access.
- Audit CI/CD pipelines. Build systems often have powerful credentials.
- Document emergency rotation procedures.
- Train developers and IT teams. Secret exposure is an architecture and process problem, not only a coding mistake.
What Small and Mid-Sized Businesses Should Review Now
Identity abuse rarely arrives through one channel. The same discipline applies to Microsoft Teams help desk impersonation and to unmanaged IoT devices like those in the XCPE botnet.
SMBs may not think of themselves as software companies, but many now rely on websites, SaaS integrations, Power Automate, Microsoft 365, Azure-hosted applications, third-party APIs, marketing platforms, payment integrations, AI tools, and developer contractors. Each one can introduce credentials that need governance.
A practical starting point is to inventory the applications and service accounts that connect to Microsoft 365, Azure, GitHub, CRM, websites, analytics, backup platforms, and other business-critical services. Then identify which integrations still depend on long-lived static credentials and which can move to safer authentication models.
What This Means for South Florida Businesses
This campaign is global, not South Florida-specific. But organizations in Miami, Miami Lakes, Fort Lauderdale, Broward County, and the wider South Florida market use the same Microsoft 365 tenants, Azure applications, GitHub repositories, SaaS platforms, outsourced development partners, and API integrations targeted by this class of attack.
For smaller organizations, one overlooked application credential can be especially significant because the same integration may connect multiple systems without the visibility or segregation found in larger enterprises.
How CompuAce Helps Businesses Reduce Cloud Credential Risk
Secrets management touches cloud architecture, Microsoft 365, identity, application integrations, source code, endpoint security, and incident response. CompuAce can help organizations review the technology foundations behind those risks, including:
- Microsoft 365 and Microsoft Entra security review
- Azure identity and access review
- Service principal and application-permission review
- Cloud access and least-privilege planning
- Source-code and exposed-secret risk review
- Credential rotation and lifecycle planning
- Managed identity and workload authentication planning
- Endpoint and account-security controls
- SMB cybersecurity monitoring and managed IT support
Learn more about CompuAce SMB Security, our IT solutions, CompuAce, and our editorial and sourcing standards.
Would You Know If One of Your Cloud Keys Was Public Right Now?
CompuAce helps businesses review cloud identities, application access, Microsoft 365 and Azure permissions, exposed credentials, and the controls that reduce the blast radius when a secret leaks.
Review Your Cloud & Identity Security
Exposed Secrets & Agentic AI FAQ
What did Microsoft observe attackers doing with agentic AI?
Microsoft observed threat actors using an agentic offensive security framework to search public GitHub repositories for exposed credentials, application secrets, and AI-related API keys, then use valid credentials for follow-on cloud activity.
How fast can an exposed secret turn into a cloud compromise?
In one Microsoft-observed case, an AI agent moved from searching GitHub for a client secret to exfiltrating data from Azure Key Vault in five minutes. This is an observed example, not a universal exploitation time.
Is deleting an exposed secret from GitHub enough?
No. Removing the secret from the visible file does not invalidate it. The credential should be revoked or rotated, and its historical use should be investigated.
What types of secrets are attackers looking for?
Examples include OAuth application client secrets, service principal credentials, API keys, storage keys, connection strings, database credentials, personal access tokens, SSH keys, and credentials for cloud and AI services.
What is a service principal in Microsoft Entra ID?
A service principal is an application identity used to authenticate programmatically to resources. Its effective risk depends on the permissions assigned to it.
How can GitHub help prevent secret leaks?
GitHub Secret Protection can detect supported credentials and push protection can block many supported secrets before they are committed or pushed. GitHub also scans public repositories for supported secret patterns.
Why are managed identities safer than long-lived application secrets in Azure?
Managed identities allow supported Azure workloads to obtain Microsoft Entra tokens without developers storing or managing a static application credential in source code.
What should a business do immediately after discovering a leaked cloud credential?
Treat the credential as compromised, revoke or rotate it, identify its permissions, review sign-in and audit logs, investigate any resources it could access, rotate downstream credentials if necessary, and reduce excessive privileges before restoring normal access.
Sources & Editorial Transparency
This article prioritizes first-party threat intelligence and product documentation. Incident claims are based on Microsoft Threat Intelligence. Preventive controls are cross-checked against GitHub, Microsoft Learn, and OWASP documentation. Sources were reviewed on November 9, 2026.
- Microsoft Security Blog — Threat actors use agentic AI tooling to find and abuse exposed secrets
- GitHub Docs — Secret security concepts
- GitHub Docs — Push protection
- GitHub Docs — Secret scanning alerts
- Microsoft Learn — Managed identities for Azure resources
- OWASP — Secrets Management Cheat Sheet
Editorial note: Threat infrastructure, indicators, product capabilities, and recommended mitigations can change quickly. Organizations should verify current Microsoft and GitHub documentation before making production security changes. This article does not claim that every exposed credential will be exploited or that every agentic AI framework behaves like the activity Microsoft observed.
By CompuAce Team — November 9, 2026
CompuAce Team
Senior IT consultants, cybersecurity specialists, and database engineers at Advanced Consulting Enterprises — operating in South Florida from our Miami Lakes office since 1984.
Read more about our editorial team →Need IT Support for Your Business?
Talk with a senior consultant from our Miami Lakes team — no pressure, no sales script.
Related Articles

The Smart Device on Your Network Could Already Be a Botnet Node
Google says the XCPE backdoor enrolled more than 1.3 million Android set top boxes into a botnet. Here is what businesses should learn about IoT inventory, network segmentation, device procurement, and supply chain risk.

Your IT Help Desk Just Messaged You on Teams. Are You Sure It’s Really Them?
Microsoft says attackers pose as IT support in Teams, gain Quick Assist access, then abuse AD CS. Here is how to reduce the risk.

Your HTTPS Certificate Can Be Legitimate and the Website Can Still Be Hijacked
Chrome says .gh, .sl and .as registry hijacks let attackers change DNS and obtain unauthorized certificates. Here is what businesses should review.