Thoroughly sanitize all training data prior to model development to reduce risk of poisoning attacks. Use IAM for authorization, and apply role-based restrictions to code repositories and cloud services.

Limit external libraries and dependencies that are used for the ML models as much as possible. Continuously rotate encryption keys that are used to access the user data from BigQuery and Cloud Storage.

Develop strict firewall rules to limit external traffic to Cloud Run instances. Integrate intrusion detection systems (IDS) for real-time anomaly detection on Pub/Sub message flows.