From Semantic Decisions to Single-Use Execution Authorization: A Vulnerability-Driven Runtime Security Study of the SVP Kernel

Abstract

This preprint presents SVP Kernel v0.6.3, an experimental runtime authorization architecture for AI-agent execution. The system cryptographically binds a semantic decision to the authorized request and treats the resulting authorization as single-use before execution. The study evaluates request/decision binding, replay resistance across process restart, authorization substitution, decision/rule/threshold mutation, forged actions, concurrent distinct authorizations, and same-authorization race behavior using synthetic execution experiments. The results demonstrate the measured behavior of the prototype under these test conditions; they do not establish production security, distributed deployment security, production key management, authorization revocation, or resistance to compromise of the prototype signing secret.