ModeLeak: Privilege Escalation to LLM Model Exfiltration in Vertex AI

Unit42 researchers found two serious vulnerabilities in Google Vertex AI: a privilege-escalation via malicious custom jobs that abuse service-agent permissions to access project resources, and a model-exfiltration (model-to-model infection) where a deployed poisoned model can obtain and download other ML and fine-tuned LLM adapter weights; the report includes detailed attack steps, impact analysis, and mitigation recommendations, and notes Google has patched the specific issues disclosed.