ARTEX AI Agents Hunt Leaked GitHub Secrets: How to Close the Five-Minute Path to Your Cloud
Microsoft saw an ARTEX AI agent go from a leaked GitHub secret to Azure Key Vault data in five minutes. How to find, rotate, and lock down app credentials.
- Author
- Ackatec
- Published
- Reading time
- 6 min read
- AI security
- Cloud security
- Secrets management
- Identity and access
- Penetration testing
Searching public code for passwords and keys is one of the oldest moves in cloud attacks. What changed this month is the speed. On November 5, 2026, Microsoft Threat Intelligence reported that attackers are using an agentic AI offensive framework called ARTEX to search public GitHub for exposed secrets, sign in with them, and work through Azure and Microsoft 365 with little human involvement. In at least one case, the agent went from scanning GitHub for secrets to exfiltrating data from Azure Key Vault in five minutes.
The short answer: treat any secret that has ever appeared in public code as compromised, rotate it now, and then shrink what each application identity can reach. Deleting the commit does not help, because a leaked credential stays usable until someone revokes or rotates it.
What is ARTEX?
ARTEX is an agentic penetration testing framework: an AI agent that is given a goal and decides for itself which tools and steps to use. According to GitGuardian's analysis (opens in a new tab), it was published on GitHub in late July 2026 and pulled on October 8, right after the first report of use by threat actors. Removing it did not stop the activity. GitGuardian cites Censys data showing roughly 1,400 publicly exposed ARTEX instances as of November 4.
Microsoft notes that ARTEX by default uses ordinary tools such as Nmap, curl, Python HTTP clients, and headless Chromium. The novelty is not a new exploit. It is an agent that can find a credential, test it, read the results, and choose the next step at machine speed.
How are AI agents abusing secrets leaked on GitHub?
Across the cases described in Microsoft's report (opens in a new tab), the pattern looks like this:
- Search. Starting in August 2026, the agent sent large volumes of GitHub queries that often combined a target organization's name with keywords for secrets, cloud services, or AI providers.
- Sign in. As early as October 28, Microsoft saw successful sign-ins as service principals, the identities applications use to reach cloud resources, across dozens of applications in multiple tenants.
- Map and harvest. The agent enumerated subscriptions, permissions, and AI model deployments, then went after operations that return more credentials: storage account keys, App Service settings, connection strings, and Kubernetes credentials. It read Key Vaults, downloaded SharePoint and OneDrive files, searched Azure DevOps for secrets, and used Run Command to execute scripts on Azure virtual machines.
- Adapt. When a cloud API call failed, the agent fixed it. In one sequence, a failed Key Vault request was rerun about 28 seconds later with a corrected API version and succeeded.
Two details deserve attention from security leaders. First, in some cases a multitenant application credential was compromised at the vendor that owned the app and then used to reach several downstream customer tenants. Your exposure can come from a supplier's leak, not just your own. Second, the agents repeatedly focused on credentials for AI services, including API keys for major AI providers.
Why does deleting the leaked secret not fix the problem?
Because the credential itself still works. Microsoft states that publicly exposed credentials remain usable until revoked or rotated, and that copies can survive in edit history, caches, archives, and logs. GitGuardian's data shows how long that tail is: it counted 28.65 million new hardcoded secrets on public GitHub in 2025, and reports that 64% of the valid secrets it found in 2022 are still active.
What should you do this week?
Start with containment, then hunt, then harden. Keep notes, because you may need them for customers, insurers, or regulators.
Contain
- Find what is already exposed. Scan your public repositories, and the public repositories of your developers and contractors where you can, for credentials tied to your domains and tenants. Include issues, gists, and old forks.
- Rotate first, investigate second. For every exposed secret, revoke or rotate it, then review its historical use. Do not wait for proof of abuse.
- Clean up affected applications fully. Microsoft advises removing all client secrets, certificates, and federated credentials on an affected app, including any an attacker may have added. Tokens already issued stay valid until they expire.
- Rotate what the app could read. If an application identity was exposed, rotate everything in every Key Vault it could access, along with credentials found in documents it could open.
Hunt
- Look for failed secrets in sign-in logs. Microsoft points to Entra errors AADSTS7000215 (invalid secret) and AADSTS7000222 (expired secret) as signs that someone holds a credential for your app. A failure followed by a success is urgent.
- Look for behavior, not just IP addresses. The attackers can change IPs and user agents easily. Microsoft suggests watching for one address authenticating to three or more service principals, or one app requesting tokens for Azure Resource Manager, Key Vault, Storage, and Microsoft Graph in a short window.
- Ask your SaaS vendors. If a vendor's multitenant app has access to your tenant, ask in writing whether its credentials were exposed and what it rotated.
How do you stop the next leaked secret from becoming a breach?
You reduce two things: how often secrets leak, and how much a leaked secret can do. Both matter, because no prevention program catches every leak.
- Block secrets at push time. GitHub's push protection (opens in a new tab) for user accounts is on by default and stops you from pushing secrets to public repositories. Repository-level push protection, which blocks secrets from reaching the repositories you protect and alerts on bypasses, requires GitHub Secret Protection and is off by default, so check whether it is actually enabled.
- Remove secrets from the design. Microsoft recommends moving applications away from secret-based authentication. Managed identities (opens in a new tab) let Azure workloads authenticate without credentials in code, and workload identity federation (opens in a new tab) removes stored secrets for supported external workloads such as GitHub Actions and Kubernetes.
- Shrink each identity's reach. Apply least privilege to service principals, set maximum lifetimes for app secrets, and restrict who can register applications in your tenant.
- Add conditional access for workloads where you can. Microsoft notes that Conditional Access for workload identities requires Workload ID Premium licensing and covers only single-tenant apps registered in your tenant, so it is a partial control, not a complete one.
The OWASP Non-Human Identities Top 10 (opens in a new tab) is a useful frame for this work. Its 2025 list names secret leakage, overprivileged identities, and long-lived secrets among the top risks, which maps closely to the ARTEX attack path.
What does this mean for AI governance?
It means AI keys are now high-value credentials and need the same discipline as cloud admin keys. If your teams have adopted AI coding tools, AI gateways, or internal agents, inventory which API keys and app identities they use, who owns them, and what data they can reach. The LiteLLM incident showed a similar lesson from the supply chain side, which we covered in how to secure AI gateways and the keys they hold. Our AI governance and readiness work starts with that kind of inventory and ownership model.
How can you test whether you are exposed?
Ask the question an attacker's agent would ask: if one application secret leaked today, what could it reach in five minutes? Answering it honestly usually requires an outside view.
- A security assessment can review how application identities, secrets, and cloud permissions are managed, and rank the gaps by business impact.
- A penetration test scoped to include your cloud environment can show what an attacker could actually reach from a single foothold, with evidence your engineers can act on.
Neither replaces rotating exposed credentials today. Do that first. Then, if you want help answering the five-minute question for your own environment, talk with Ackatec about scoping an assessment or test.
