What is the main purpose of defining access tiers (for example: browse, inference-only, download) on an internal model hub?
- A.To ensure only users with a legitimate need can pull full model weights, while broader audiences get less privileged access like inference or metadata browsing
- B.To make the hub's user interface load faster
- C.To increase GPU utilization across the cluster
- D.To guarantee models never produce biased outputs
Why A is correct
Access tiers apply least privilege to model artifacts: most users may only need to browse model cards or call a hosted inference endpoint, while downloading full weights, which enables uncontrolled copying and exfiltration, is restricted to roles with a documented need. This mirrors NIST SP 800-53 AC-6 least privilege applied to ML assets. Tiering has nothing to do with UI speed, GPU utilization, or output bias.
Know someone studying for AI Security Fundamentals? Send them this one.