TBIVL: A Unified Identity Verification Layer for Secure Agent-to-Service Context Exchange

Abstract

Secure agent-to-service context exchange requires strong guarantees that both the calling agent and the target service are authentic. Without such guarantees, systems remain vulnerable to impersonation, context interception, and privilege abuse. While prior work has primarily focused on prompt injection defenses, output filtering, and access control policies, the verification of entity authenticity at the moment of context exchange remains underexplored. Moreover, existing transport-layer mechanisms such as mutual TLS do not sufficiently bind application-layer requests to verified agent and service identities. In this paper, we introduce the Trust-Bounded Identity Verification Layer (TBIVL), a unified application-layer framework that enforces identity verification and request-scoped authorization prior to context exchange. TBIVL integrates agent attestation, service authenticity validation, and fine-grained authorization into a single decision layer, ensuring that only trusted interactions are permitted. We implement a prototype and evaluate it through both controlled experiments and large-scale simulations. In a controlled setting (five scenarios, 1,000 requests) and a broader simulation (100 agents, 50 services, including 10 spoofed entities, across 10,000 requests), TBIVL successfully rejected all malicious requests. Additional sensitivity analysis across varying malicious traffic ratios (5–50%) showed consistent resilience with no successful attacks observed. The system maintained a mean latency overhead of less than 0.04 ms per request. These results demonstrate that combining agent attestation, service validation, and scoped authorization at the application layer can effectively mitigate impersonation and spoofing attacks while introducing minimal performance overhead.