Blog

Attackers using agentic AI exfiltrated Azure Key Vault data within five minutes in one case

Microsoft reports that agentic AI found leaked credentials in GitHub and exfiltrated Azure Key Vault data within five minutes in one case.

· By

A closed metal vault panel in a rack beside a bench with an abandoned bunch of keys and a paper log, while a hand reaches toward a red emergency-stop button.
Microsoft reports that agentic AI found leaked credentials and in one case exfiltrated Azure Key Vault data within five minutes.Image: IamVera.ai — original editorial illustration

Microsoft Threat Intelligence reports that attackers using agentic AI found leaked credentials in public GitHub repositories and, in one case, exfiltrated Azure Key Vault data within five minutes. So bound autonomous tool use outside the model with short-lived workload identities, deterministic parameter checking, credential rotation and an independently tested stop path, because prompt instructions alone are not a security boundary.

On 5 November 2026 Microsoft Threat Intelligence described an attack chain in which agentic tooling autonomously searched public repositories for exposed secrets and API keys, accessed cloud resources with valid application credentials, adjusted failed API requests and continued through to exfiltration. The observed step from discovery to exfiltration from Azure Key Vault took five minutes in that one case. Microsoft advises rotating or revoking exposed credentials immediately, giving service principals minimal rights, monitoring workload identities and examining API and audit logs.

What sets this event apart from an ordinary credential leak is not the leak itself, but the pace and the adaptation. That shifts the risk from who holds the secret to what the identity behind that secret can reach before anyone intervenes.

What exactly did Microsoft establish about agentic tooling and leaked credentials?

Microsoft Threat Intelligence established that an agentic framework scanned public GitHub repositories for secrets, accessed cloud resources with the credentials it found, adjusted API requests after errors and, in at least one case, went from discovery to Key Vault exfiltration within five minutes. Microsoft stresses that those five minutes are an incident-specific observation, not a universal attack speed.

The attack chain is recognisable in its separate parts: an exposed secret, a valid workload identity, reconnaissance of available APIs, adjusted attempts after refusals, access across multiple services and finally data movement. What is new, in our analysis, is that those parts were strung together by agentic tooling with substantial autonomy; the available source does not establish that no human operator was involved at every step.

Why is a prompt instruction not a sufficient security boundary for tool use?

An instruction to the model describes desired behaviour, but it does not determine which APIs are reachable, which parameters are accepted or which identity executes the call. Whoever relies on that lets the security boundary be set by text that the agent itself can reinterpret or that an attacker can influence through poisoned data.

Microsoft Learn addresses precisely this distinction in its guidance on reducing agentic risk: separate instructions from data and parameters, limit the reachable actions outside the model and retain independent monitoring and interruption. The OWASP GenAI Security Project names, in its Top 10 for agentic applications for 2026, risks such as agent hijacking, data leaks, supply-chain compromise and agent sprawl as systemic problems of systems that plan, act and decide. In our assessment, the Microsoft incident shows why both sources place the boundary outside the prompt layer: the agent adjusted its API requests, so a textual instruction would not have bounded the behaviour.

Which controls enforce autonomous API use outside the model?

The usable boundary lies in the infrastructure around the model, not in the prompt. Microsoft Learn, OWASP and CrowdStrike describe controls that together form a chain: identity, authorisation, validation, oversight and interruption. Below is the core, translated into what an organisation enforces itself.

  • Unique, auditable agent identity. Give each agent its own auditable identity with task-specific, short-lived rights, rather than reusing a broad service principal.
  • Secrets outside the model context. Keep credentials away from what the model sees, so that a hijacked agent cannot read them out or pass them on.
  • Allowlist of tools and resources. Permit only pre-approved APIs and destinations; anything beyond that is deterministically blocked.
  • Deterministic parameter validation. Check the scope and parameters of every call outside the model, before execution, so that requests adjusted after a refusal do not succeed anyway.
  • Credential rotation and revocation. Rotate or revoke exposed secrets immediately, as Microsoft advises, and test that revocation actually cuts off access.
  • Human approval for irreversible actions. Require sign-off for high-impact or non-reversible calls.
  • Behavioural oversight across identity, host and service. Correlate signals; CrowdStrike describes baselining and policy-drift detection, though it marks some features as beta or subject to change.
  • Independently tested stop and containment path. Ensure an agent can be interrupted and access can be revoked in a coordinated way, and test that under realistic failure scenarios.

CrowdStrike’s account of detecting and stopping agent insider threats, published on its blog on agent insider threats, describes agents as non-human identities that may operate under an employee's identity and may drift after a tool or instruction is manipulated. That aligns with runtime authorisation that curtails an agent's authority when reliability gradually declines, and with separation of duties with dual authorisation for AI agents, where the agent, the policy layer and the human approver remain separated.

What does this mean now for directors, lawyers and CISOs?

Our analysis: because the agent observed by Microsoft adjusted its API requests after refusals, an instruction to the model offers no protection against escalation; so the CISO must set up the tool allowlist and the deterministic parameter validation outside the prompt layer before the next production release. Because Microsoft observed a path from discovery to Key Vault exfiltration of five minutes in one case, an approval process that takes hours may be too slow, so the stop and revocation path must work in minutes; have that tested under a realistic failure scenario. Because the agent acted with a valid workload identity, the director must have it recorded which service principal feeds which agent, with which rights, and reduce those rights to the strictly necessary. Because Microsoft advises immediate rotation or revocation of exposed credentials, a rotation policy without tested revocation is false security; so the responsible manager must demonstrate this quarter that a revoked secret really no longer grants access. The practical translation, as a checklist:

  • Inventory which agents run under which workload identities and prune excess rights.
  • Move the tool allowlist and parameter validation to a layer outside the model.
  • Test revocation and interruption for actual operation and for speed.
  • Record per call: agent identity, delegated context, tool, parameters, authorisation decision, API response and data movement.

More depth is available in our topic hub on agentic AI and AI agents, in a dedicated identity chain for autonomous AI agents and in the OWASP risks for language models translated into controls.

Sources and references

  1. Threat actors use agentic AI tooling to find and abuse exposed secretsMicrosoft Threat Intelligence · 2026-11-05
  2. Reduce autonomous agentic AI riskMicrosoft Learn · 2026-11-04
  3. OWASP Top 10 for Agentic Applications for 2026OWASP GenAI Security Project · 2025-12-09
  4. CrowdStrike Extends the Agentic SOC to Detect and Stop Agent Insider ThreatsCrowdStrike · 2026-11-04

Sources: The article draws on Microsoft Threat Intelligence, Microsoft Learn, the OWASP GenAI Security Project and CrowdStrike.

← All articles in this topic ← All articles