Threat actors use agentic AI tooling to find and abuse exposed secrets

Source: Microsoft Security Blog

Author: Microsoft Threat Intelligence and Microsoft Security Research

URL: https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/

ONE SENTENCE SUMMARY:

Threat actors use agentic AI to rapidly find exposed GitHub secrets, abuse Entra service principals, and exfiltrate Azure data.

MAIN POINTS:

1. Agentic AI frameworks scan public GitHub at scale for credentials and API keys.
2. ARTEX reduces hands-on-keyboard time by chaining discovery, analysis, and exploitation automatically.
3. Microsoft observed GitHub-to-Azure Key Vault exfiltration completed in approximately five minutes.
4. Attackers misused leaked application credentials to sign into Entra ID as service principals.
5. A compromised multitenant vendor app credential enabled access across multiple downstream tenants.
6. Campaign infrastructure included sign-ins from 199.19.77.84 and 199.19.77.18 (Tor exit node).
7. Post-auth activity targeted Key Vault secrets, Storage data, App Service tokens, and AI resources.
8. Adaptive cloud API behavior showed rapid retries with corrected API versions to obtain HTTP 200.
9. OAuth app abuse enabled Azure discovery, Azure DevOps secret searching, and VM Run Command execution.
10. Microsoft provided detections across Defender/Sentinel plus KQL hunting for Entra, Graph, and cloud logs.

TAKEAWAYS:

1. Treat any publicly exposed secret as compromised; rotation is mandatory even after deletion.
2. Block known malicious IPs via Conditional Access named locations, targeting workload identities carefully.
3. Prefer managed identities over long-lived OAuth client secrets to eliminate code-based secret exposure.
4. Hunt behaviorally for rapid multi-resource token use and many service principals per single IP.
5. Use deception (honeypots, tarpits, fake wins) to divert autonomous agents and create high-confidence signals.