When AI infrastructure becomes the target: Securing gateways and control points

Microsoft observed a campaign targeting AI control-plane components (LiteLLM gateway, RAGFlow retrieval, Kestra orchestration) where attackers leveraged exposed application surfaces and public CVEs to gain execution, harvest model-provider keys and database credentials, persist inside runtimes, and monetize compute via XMRig cryptomining. The report details multi-stage attack chains per product, includes IOCs (IPs, domains, file hashes), MITRE technique mappings, advanced-hunting queries, and concrete mitigation recommendations to protect AI gateways, constrain egress, and monitor for gateway-originated execution and secret access.