Google AI Platform Bugs Leak Proprietary Enterprise LLMs

Palo Alto Networks Unit 42 disclosed two vulnerabilities in Google Vertex AI that together allowed privilege escalation via custom jobs/service agent permissions and deployment of a poisoned model capable of exfiltrating all fine-tuned ML and LLM models in a project; Google has applied fixes. The research highlights risks of model-to-model infection, excessive deployment permissions, and the need to separate and vet development/test models from production to prevent unauthorized access and proprietary model theft.