perishable studies the security model for AI API access in an agent-driven world. It implements scoped, time-limited token proxies that provide secure LLM access without exposing long-lived credentials — exploring zero-trust patterns for AI infrastructure.