# Microsoft: An AI Agent Went From a Leaked GitHub Secret to Azure Key Vault Data in Five Minutes

_Microsoft Threat Intelligence reported on November 5, 2026, that attackers point AI agents at public GitHub repositories to find exposed app secrets and act on them, using frameworks such as ARTEX. In one case an agent went from scanning GitHub to taking Azure Key Vault data in five minutes._

By Johannes Keienburg, CEO & Founder  
Published: 2026-11-12  
Source: https://www.cakewalk.security/blog/microsoft-ai-agents-github-leaked-secrets

---

## AI Agents Found Leaked App Secrets on GitHub and Used Them Against Their Owners

Microsoft Threat Intelligence has watched attackers point AI agents at public GitHub repositories since August 2026, looking for the secrets applications use to sign in to cloud services. In a report published on November 5, 2026, Microsoft ties the activity to agentic penetration testing frameworks such as ARTEX. The agents run large volumes of GitHub searches that pair a target organization's name with keywords for secrets and cloud services. In at least one case, an agent went from scanning GitHub to taking data in five minutes. The data sat in Azure Key Vault, Microsoft's service for storing keys and secrets. A leaked secret lets whoever holds it sign in as the application, a [non-human identity](/glossary/non-human-identity) with whatever access the application holds.

> Agentic tooling like ARTEX allows threat actors to not only scan for secrets across multiple repositories at a quicker pace but also use the agents to analyze the output and conduct follow-on attacks, cutting down on the need for the threat actor to have hands-on-keyboard interactions throughout the attack chain.

_Microsoft Threat Intelligence, Threat actors use agentic AI tooling to find and abuse exposed secrets, November 5, 2026_

From October 28, Microsoft saw an agentic framework use leaked or stolen application credentials on dozens of applications across several organizations' cloud directories. It signed in as [service principals](/glossary/service-account), the identities applications use to authenticate. In some cases the credential belonged to a vendor-owned application used by many customer organizations. It was compromised at the vendor and then used against several customers, as in the [Qubole breach on AWS](/blog/qubole-cross-account-roles-aws-breach) earlier this month. The attacker went after operations that return further credentials, such as storage account keys and connection strings, while the compromised applications downloaded files from SharePoint and OneDrive. In three logged cases, the agent reran a failed cloud API request with a different API version about 28 to 80 seconds later and got a successful response. Microsoft says the agents repeatedly focused on credentials for AI services.

## Microsoft Says a Leaked Secret Still Works After It Is Deleted From Code

Microsoft tells organizations to scan their code for secrets, treat any exposed secret as compromised and rotate it immediately. It recommends moving [OAuth](/glossary/oauth) applications away from secret-based authentication to managed identities, applying [least privilege](/glossary/least-privilege) to service principals and restricting who may register new applications in a tenant.

> Publicly exposed credentials remain usable until revoked or rotated; removing the original disclosure alone does not remediate the exposure.

_Microsoft Threat Intelligence, Threat actors use agentic AI tooling to find and abuse exposed secrets, November 5, 2026_

Microsoft adds that blocking the default user-agent strings these frameworks send will not stop an agent that changes them quickly, though spotting them can alert defenders.

> A client secret in a public repository signs in as its application for whoever finds it, with no person behind the sign-in to answer for it. None of these sign-ins passed through Cakewalk's MCP Gateway, so no Cakewalk check sat in their path. For the calls the gateway does govern, it keeps the upstream credential encrypted on its side and adds it at call time, so that credential never sits in the agent's code or config. Each agent works for the person who connected it, with access kept to what that work needs. Before a governed call runs, the gateway weighs it against policy and records which rule allowed or denied it.

_Comment from Gil Röder, CPO & Co-Founder_

Source: Microsoft Threat Intelligence, [Threat actors use agentic AI tooling to find and abuse exposed secrets](https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/), November 5, 2026.
