A Symbolic Analysis of Agent Evidence Packages
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
This article studies an Agent Evidence Package (AEP) profile from the Agent Trust Framework (EATF) and a Model Context Protocol (MCP) attestation profile using symbolic protocol modeling in Tamarin. The starting point is an ordinary engineering question: if a valid signed evidence package is shown to a verifier twice, is the second acceptance only another inspection, or a new action? The model represents agent authorization, gateway forwarding, signer issuance, timestamping, ledger anchoring, and verifier acceptance under explicit Dolev--Yao and trust-anchor assumptions. Several correspondence-style properties verify, including signer and agent authentication, payload integrity, timestamp-event linkage, forward-integrity-style pre-compromise attribution, and resistance to a malicious gateway acting as a signing oracle. The clearest negative result concerns replay semantics: a stateless verifier can accept the same otherwise valid package hash more than once, so freshness and ledger anchoring alone do not imply unique verifier acceptance. A replay-cache variant verifies uniqueness under added verifier state. The contribution is model-bounded and profile-dependent: replay resistance must be stated as a verifier-state or challenge-bound property when a downstream application treats acceptance as a fresh event.