Internet-Draft AI Agent Network Trust June 2026
Yu Expires 31 December 2026 [Page]
Workgroup:
Internet Engineering Task Force
Internet-Draft:
draft-yu-network-trust-framework-for-ai-agents-00
Published:
Intended Status:
Informational
Expires:
Author:
H. Yu
China Internet Network Information Center

Network Trust Framework for AI Agents

Abstract

AI agents are expected to communicate across platforms, organizations, clouds, edge environments, and administrative domains. Existing agent mechanisms mainly focus on application-layer discovery, capability description, identity, authorization, tool invocation, and agent-to-agent collaboration. These mechanisms are important, but they often treat the IP network as a transparent connectivity substrate.

This document describes a framework for using network-layer and network-control capabilities to enhance trust for AI agent communication after agent discovery. The framework introduces an Agent Interconnection Hub, trusted transport tags, policy-driven path selection, and audit indexing as building blocks for recognizable identity, accountable paths, enforceable policy, and traceable invocation records.

This document does not define a new agent discovery protocol, a new application-layer agent collaboration protocol, a new authentication or authorization mechanism, a new IPv6 extension header, or a new SRv6 behavior. It provides architectural considerations and requirements for future work.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 31 December 2026.

Table of Contents

1. Conventions Used in This Document

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

2. Introduction

AI agents are moving from isolated application components toward large-scale, cross-domain collaboration. An agent may discover other agents, invoke tools, access data, call APIs, and perform tasks on behalf of a user, organization, or system. When these actions become continuous, autonomous, and cross-domain, agent communication becomes not only an application-layer issue but also an Internet infrastructure issue.

Application-layer mechanisms such as agent cards, tool metadata, directory services, DNS-based discovery, decentralized identifiers, certificates, authorization tokens, and agent collaboration protocols can help an agent find a peer and determine what it can do. However, after discovery, the communication still traverses networks that may only observe IP addresses, transport connections, and limited metadata. The network may not be able to associate a flow with the responsible agent, the delegated authority, the task risk, the intended path, or the audit obligation.

This document explores how network-layer and network-control capabilities can enhance trust for agent communication without replacing existing application-layer mechanisms. The goal is to make discovered agents not only reachable, but also more trustworthy, path-accountable, policy-enforceable, and auditable.

3. Scope and Non-Goals

This document focuses on a framework for network-layer trust enhancement for AI agent communication. It discusses how identity verification results, task risk, network policy, path selection, and audit indexing may be associated after agent discovery.

This document discusses:

This document does not define:

4. Terminology

Agent:
A software entity that can perceive context, reason, plan, invoke tools, communicate with other agents, and perform tasks on behalf of a user, organization, or system.
Agent Communication:
Communication initiated by, received by, or mediated for an agent, including tool calls, API calls, data access, agent-to-agent calls, and gateway-mediated interactions.
Agent Interconnection Hub:
A trusted mediation point between application-layer discovery systems and the network. It performs discovery adaptation, trust verification, network policy generation, path selection, and audit indexing.
Trusted Transport Tag:
A signed or integrity-protected structure associated with an agent communication task. It carries minimum necessary metadata such as identity digests, task type, risk level, network service requirement, audit level, validity period, and revocation pointer.
Network-Layer Trust Enhancement:
The use of network-layer or network-control capabilities, such as locator binding, policy tags, path selection, service chaining, verification records, and audit indexes, to improve the trustworthiness, controllability, and traceability of agent communication.
Agent Domain:
An administrative or operational domain in which agents, gateways, registries, policy systems, audit systems, or network functions are managed under a common authority.

5. Motivation

Current agent interconnection mechanisms are largely developed at the application layer. They can describe capabilities, publish endpoints, exchange credentials, call tools, and coordinate agent-to-agent tasks. These functions are necessary, but they do not by themselves provide all of the network-layer support required by high-value or high-sensitivity agent communication.

Three gaps are particularly relevant.

These gaps make it difficult to answer operational questions such as which agent initiated a call, under which authority, for what type of task, through which path, under which network policy, and with what verification result. A network-layer trust enhancement framework attempts to provide a coherent way to relate these questions without requiring the network to inspect application payloads.

6. Design Goals

Network-layer trust enhancement for agent communication has the following design goals.

Recognizable identity:
The network-control environment can associate a communication task with identity digests, authorization status, and trust state without exposing unnecessary identity details.
Accountable paths:
The network-control environment can select paths according to task risk, data sensitivity, latency, jurisdiction, and audit requirements.
Enforceable policy:
Application-layer task context can be translated into network-executable policy for gateways, service chains, path controllers, or security functions.
Traceable audit:
Source-side, destination-side, and necessary intermediate nodes can produce audit records that share a common index.
Privacy-aware operation:
Network nodes receive only the information needed for routing, policy matching, enforcement, and audit indexing.
Incremental deployment:
The framework reuses existing Internet, IPv6, SRv6, certificate, credential, gateway, and audit mechanisms where possible.

7. Architectural Components

This framework is organized into four logical planes. A deployment may combine these functions in a single gateway, distribute them across multiple nodes, or integrate them with existing operator, enterprise, or industry infrastructure.

7.1. Discovery Adaptation Plane

The discovery adaptation plane interfaces with existing discovery and capability-description mechanisms, such as DNS records, service-binding metadata, agent cards, MCP server metadata, DID documents, industry registries, enterprise catalogs, or platform directories.

This plane normalizes discovery results into a common access description that can include the target entry point, supported capabilities, access conditions, security requirements, and minimum information needed by the trust-verification and network-control planes.

7.2. Trust Verification Plane

The trust verification plane verifies agent identity, responsible entity, authorization, credential state, and task risk. It may use agent identity credentials, domain certificates, DID/VC material, platform signatures, enterprise authorization, API tokens, gateway certificates, and session signatures.

RPKI and ROA can be used to validate the origin of an IPv6 prefix [RFC6480], but prefix-origin validation alone does not prove agent identity. Agent trust requires additional certificate, credential, authorization, signature, or hub-mediated validation.

7.3. Network Control Plane

The network control plane translates verified agent context and task risk into executable network policy. It may use IPv6 reachability [RFC8200], SRv6 path steering [RFC8986], gateway selection, service chaining, application-aware networking, security inspection points, or network-management interfaces.

Example policies include best-effort forwarding for low-risk tasks, trusted paths for government or financial tasks, mandatory security inspection for sensitive data access, audit-node traversal for cross-domain tasks, and explicit low-latency paths for time-sensitive operations.

7.4. Audit Governance Plane

The audit governance plane creates and correlates audit records for agent communication. It records structured evidence such as caller and callee identity digests, task type, policy identifier, trusted transport tag digest, verification result, path policy, important transit nodes, timestamp, and exception state.

The audit governance plane is not intended to record application plaintext by default. Recording prompts, tool inputs, tool outputs, or user data can significantly increase privacy risk and needs separate authorization and legal basis where applicable.

8. Trusted Transport Tags

A trusted transport tag is a framework concept for associating an agent communication task with trust and policy metadata. This document does not specify a wire format. Future work could define a format, or deployments could use existing token, certificate, attestation, gateway, or controller mechanisms.

A trusted transport tag can include:

A trusted transport tag should be designed for minimization. It should not directly expose full user identity, full organization identity, prompts, business content, or sensitive credentials to arbitrary intermediate nodes.

9. Example Communication Flow

A typical flow using the framework can be described as follows.

  1. The caller agent discovers a target agent through DNS, an agent directory, an agent card, MCP metadata, an industry registry, or an enterprise catalog.
  2. The caller sends the request through a local Agent Interconnection Hub.
  3. The discovery adaptation plane parses the discovery result and extracts the minimum information needed for policy processing.
  4. The trust verification plane verifies the caller, callee, authorization state, credential status, and task risk.
  5. The hub creates a trusted transport tag or refers to an existing policy index.
  6. The network control plane selects an appropriate path, gateway, service chain, or SRv6 policy according to task requirements.
  7. The destination-side hub verifies the tag, signature, authorization, and policy compatibility before delivering the request to the target agent or service.
  8. The source-side hub, destination-side hub, and any required intermediate trust nodes create audit records using a common audit index.
  9. If verification fails or risk changes, the system can revoke the tag, block the request, reroute the flow, reduce privileges, or generate an alert.

This flow is illustrative. It does not require every deployment to use the same hub topology or the same path-control mechanism.

10. Deployment Considerations

The framework is intended for incremental and distributed deployment. It is not necessary or desirable to require all public Internet agent traffic to pass through a single mandatory control point. Initial deployment is more likely in high-value or high-sensitivity environments where identity, authorization, path control, audit, and accountability requirements are explicit.

Candidate deployment environments include:

Agent Interconnection Hubs can be deployed as enterprise gateways, operator edge nodes, industry trust nodes, regional exchange points, or cross-domain mediation nodes. Different deployments may implement different subsets of the logical planes described in this document.

11. Open Issues for Future Work

This document intentionally leaves several topics for future work.

12. Security Considerations

Agent communication can involve delegated authority, tool invocation, data retrieval, cross-domain access, and autonomous decision-making. Weak linkage among agent identity, network flows, authorization decisions, and audit evidence can make impersonation, misuse, excessive privilege, and incident reconstruction more difficult to manage.

A trusted transport tag or equivalent policy object can itself become an attack target. Implementations need to consider integrity protection, authentication of tag issuers, replay protection, scope limitation, expiration, revocation, downgrade attacks, substitution attacks, and confused-deputy risks.

Agent Interconnection Hubs are security-sensitive components. If a hub is compromised, an attacker might issue false trust metadata, suppress audit records, reroute flows, or authorize unintended communication. Deployments need operational controls, key management, monitoring, least-privilege administration, and fail-safe behavior.

This document does not define a new protocol mechanism. Any future protocol work derived from this framework will need protocol-specific security analysis.

13. Privacy Considerations

Agent communication can reveal sensitive information about users, organizations, tasks, data sources, tools, models, policies, and business relationships. Trust metadata and audit records can also become sensitive, even when application content is encrypted.

Deployments should minimize what is exposed to network nodes. Full identity, business content, prompts, tool inputs, tool outputs, and detailed reasoning traces should not be included in network-visible metadata by default. Identity digests, policy classes, short-lived tags, and audit indexes can reduce exposure, but they can still be linkable if reused too broadly.

Audit designs need to balance traceability and minimization. More complete logs can improve accountability, but can also create new data stores and new privacy risks. Access to audit records should be controlled and logged.

14. IANA Considerations

This document does not request any IANA action.

15. Normative References

[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/info/rfc8174>.

16. Informative References

[RFC8200]
Deering, S. and R. Hinden, "Internet Protocol, Version 6 (IPv6) Specification", STD 86, RFC 8200, DOI 10.17487/RFC8200, , <https://www.rfc-editor.org/info/rfc8200>.
[RFC8986]
Filsfils, C., Ed., Camarillo, P., Ed., Leddy, J., Voyer, D., Matsushima, S., and Z. Li, "Segment Routing over IPv6 (SRv6) Network Programming", RFC 8986, DOI 10.17487/RFC8986, , <https://www.rfc-editor.org/info/rfc8986>.
[RFC6480]
Lepinski, M. and S. Kent, "An Infrastructure to Support Secure Internet Routing", RFC 6480, DOI 10.17487/RFC6480, , <https://www.rfc-editor.org/info/rfc6480>.
[I-D.yu-ai-agent-ipv6-networking-considerations]
Yu, H., "IPv6 Networking Considerations for AI Agent Communication", <https://datatracker.ietf.org/doc/draft-yu-ai-agent-ipv6-networking-considerations/>.
[I-D.pioli-agent-discovery]
Pioli, L., "Agent Registration and Discovery Protocol", <https://datatracker.ietf.org/doc/draft-pioli-agent-discovery/>.
[Agent2Agent-Archive]
IETF, "IETF Agent2Agent Mailing List Archive", <https://mailarchive.ietf.org/arch/browse/agent2agent/>.

Author's Address

Haisheng Yu
China Internet Network Information Center