Inside 90 days of attacks on AI infrastructure

Wiz Threat Research observed sustained active attacks against self-hosted and managed AI infrastructure over 90 days, detailing three core patterns — MCP gateway/server exploitation (leading to unauthenticated RCE via LiteLLM CVEs), blind prompt-injection against agent frameworks to trigger out-of-band callbacks and fetch payloads, and AI-native post-exploitation focused on in-memory credential harvesting and stealthy cryptominer deployment — and provides IOCs and defensive recommendations for inventory, authentication, egress restriction, runtime monitoring, and urgent patching.