DNS for AI Discovery

Internet-Draft	DNS-AID	November 2026
Mozley, et al.	Expires 6 May 2027	[Page]

Abstract

This document specifies DNS for AI Discovery (DNS-AID), a method for publishing AI agents in the Domain Name System (DNS) so that other agents can discover them. It defines three discovery use cases, ordered by increasing latency and computational cost. In the known-agent case, the requestor knows both the agent and its domain, and a single DNS query returns connectivity information and key metadata. In the known-organization case, the requestor knows the domain but not which agent provides a required capability, and queries a well-known organizational index to get connectivity information to enable a capability search of a catalog, agent registry or database. In the known-capability case, the requestor knows only the required capability and must rely on an external directory or search service. This document addresses the known-agent and known-organization cases. The known-capability case is out of scope, but search and cataloging services can be built from the organizational index defined here. Once a client has learned an organization’s agents, later interactions can use the known-agent case, relying on a minimal, cacheable record set that each organization controls. DNS-AID uses Service Binding (SVCB) records, optionally with DNS-Based Service Discovery (DNS-SD) labels, DNSSEC, and DANE TLSA records. This gives agents direct connections that no third party mediates and works the same way on public and private networks. No new resource record types, opcodes, or response codes are introduced.¶

1. Introduction

This document standardizes an approach for publishing AI agents in the DNS so that other agents can discover them. Discovery can be initiated based on one of three use cases, where they are ranked in ascending order of latency and computational power required to complete:¶

Known Agent - The requestor knows both the agent and its origin domain, and a single DNS query provides connectivity information and key metadata (see Section 3.1)¶
Known Organisation - The requestor knows the domain but not which agent provides a required capability, and an organization-level index returns connectivity information to a catalog, agent registry or database from which one can be selected (see Section 3.2)¶
Known Capability - The requestor knows the required capability, but not the agent or domain, or knows neither, in which case it must rely on an external directory or search service (see Section 3.3)¶

Of the above discovery use cases, (1) and (2) are addressed by this document. The third use case can be derived from (2) based on a well-known entry point to an organization's index of agents, such that any organization could build a search service based on this information. In the event the capability is unknown, external search is likely used anyway, which is treated like (3).¶

1.1. Design

DNS-AID is designed so that, once a client has learned an organization's agent capabilities, subsequent queries can use the first use case. The intention is to provide a minimum DNS record set, that is cacheable or even learnable as an agentic skill, with each organization able to control the advertisement of its own agents. This also provides consumers of agent services with a direct connection method for agentic workloads not mediated by a third party. Organizations can use the same approach within internal networks, providing consistency across public and private networks. The mechanism uses SVCB and DNS-SD to include connectivity and metadata within a single DNS record, and optionally utilizes DANE TLSA records to support efficiently establishing encrypted communication.¶

A deployment using the DNS-AID approach uses the following:¶

Service Binding (SVCB) records [RFC9460] for the agent endpoint and protocol¶
TLSA records [RFC6698] MAY be used for the agent's TLS identity¶
The records SHOULD be DNSSEC-signed [RFC9364] for data origin authentication and data integrity, if TLSA records are used they MUST be signed¶
DNS-Based Service Discovery (DNS-SD) [RFC6763] MAY be used for compatibility with existing service-discovery tooling¶
Several extensions - agent-to-zone control proofs, query-time policy signaling, and cross-domain search - are being explored in parallel and are summarized under Section 5.¶

Section Section 4 defines TXT records as the fallback where SVCB, or the SvcParamKeys below, cannot be published or served.¶

1.2. Determinism

DNS discovery is deterministic and cacheable: a given query against a given zone returns the same SVCB and TLSA RRsets, signed and time-bounded, until the publisher rotates them. The downstream agent's decisions over the discovered metadata are not deterministic. Implementations and reviewers SHOULD treat DNS-AID as providing a verifiable transport for agent metadata, not as a guarantee that an agent will behave predictably given that metadata.¶

3. Discovery Use Cases

This section of the document provides details of the three generic use cases, including DNS record usage.¶

3.1. Known Agent

This section covers the use case where a requestor knows both the agent and its origin domain.¶

A single SVCB record type query for agent-name.example.com returns IP addresses, transport, agent protocol, and descriptor metadata. A TLSA record type query authenticates the TLS endpoint, although implementations may use other mechanisms. This is the form publishers SHOULD support and the form requestors MUST try first.¶

The SVCB record of an agent provides connectivity and service parameters [I-D.blwm-dawn-common-discovery-parameters] to a discovering agent. These are summarized below:¶

Table 1: Connectivity and Discovery Parameters
Parameter	Meaning	Reference
`TargetName`	Target of the SVCB record, which can be the same as the agent domain name or a different FQDN	Note 1
`ipv4hint`	TargetName IPv4 addressing	Note 1
`ipv6hint`	TargetName IPv6 addressing	Note 1
`alpn`	A transport protocol suite used by the target (e.g. `alpn=h2,h3`)	Note 1
`proto`	The agent protocol (e.g. `mcp`, `a2a`)	Note 2
`proto-ver`	The agent protocol version (e.g. `1.0`, `2025-11-25`)	Note 2
`well-known`	The .well-known suffix(es)	Note 3
`path`	Endpoint path	Note 2
`desc-ref`	Descriptor reference	Note 2
`desc-ref-digest`	Descriptor reference digest	Note 2

Notes:¶

Service Binding and Parameter Specification via the DNS (SVCB and HTTPS Resource Records) [RFC9460]¶
Common Agent DNS Discovery Parameters [I-D.draft-blwm-dawn-common-discovery-parameters]¶
An SVCB Service Parameter for Well-Known URI Paths [I-D.draft-mk-dnsop-svcb-well-known]¶

Operators MAY publish the same agent under an _agents.example.com inventory leaf (for organizations that prefer a single containing prefix for their agents) or under a canonical DNS-SD label (so off-the-shelf DNS-SD clients can enumerate). In these cases SVCB AliasMode MUST be used to point at the primary owner.¶

Not all parameters are needed in all cases, for instance a record for an agent protocol with a specification including a single fixed .well-known suffix as a description of version and capability (e.g. an agent card) may only need alpn and proto. Where an agent supports multiple protocols, or a single protocol has custom protocol bindings to support different transport mechanisms or communication patterns, then additional parameters will signal this to a discovering agent and prevent the need for probing for capability.¶

Examples of these records are shown below in Figure 1 and their usage in Figure 2:¶

Figure 1: SVCB Record Examples

Figure 2: Discovery Example 1: Known Agent Discovery

3.1.1. Agents Supporting Multiple Protocols

n agent that supports more than one agent-layer protocol publishes one SVCB record per protocol. Each record carries proto with a single protocol value and alpn with the HTTP transport versions. A client processes the RRset and selects the record matching the protocol it wants.¶

Each agent protocol MUST be a separate record in the RRset:¶

Multiple agent protocols MUST NOT be combined in a single proto value:¶

An agent hosted at a service provider uses a distinct TargetName for each protocol endpoint:¶

Figure 3: SVCB Records for Multi-Protocol Support

3.2. Known Organization

This section covers the use case where a requestor knows the organizational domain, but not which agent provides a required capability. This is likely to be an agent searching for a capability from an organizational database such as [ARD-SPECIFICATION], [I-D.draft-jimenez-agent-directory] or other agent directory.¶

An SVCB record type query for _index._agents.example.com returns a pointer to an organization-specific registry of all agents, which is out of the scope of this draft. The requestor uses the data returned to select an agent and cache the selection so that the process in Section 3.1 can be used for subsequent interaction. The data provided at _index._agents.{domain}, protocols and schemas are out of scope for DNS-AID.¶

The TargetName MUST be used and MUST not contain underscores as used in DNS-SD labels, as public x.509 certificates will be used in communications with the index. Following the _index._agents.example.com SVCB record query, agents will use the TargetName for any TLSA record query.¶

The labels _index._agents have been chosen as DNS-SD labels with the intent of registering them with IANA, as it is not possible to register a generic name such as agent-index.¶

An index record SHOULD carry the parameters a client needs to reach and cache the index such as alpn and ipv4hint for connectivity information. Then parameters as applicable to the catalog, agent registry or database implementation such as well-known or path for the HTTP endpoint, desc-ref for the descriptor of the service, and proto where the index is itself served over an agent protocol.¶

Examples of these records are shown below in Figure 4:¶

Figure 4: Organizational Index Record Examples

Figure 5: Discovery Status 2: Index-Based Discovery with Service Selection

3.3. Known Capability

This section covers the use case where the requestor knows the required capability, but not the organization or agent.¶

This is a search based process and is out of the scope for DNS-AID. Directory services that aggregate records across organizations could be derived from Section 3.2 by querying _index._agents.{domain} and using that data for a search service.¶

5. Future Work and Experimental Mechanisms

This section collects mechanisms that have been prototyped against the DNS-AID record set but that are not yet ready for normative specification in this document. Implementations MAY experiment with them; future revisions or separate documents are expected to standardize the ones that prove out. The headings below replace the briefer "Future Work & Unaddressed Portions" section in earlier revisions and consolidate the open questions raised there.¶

5.1. Agent Protocol and Version

The proto parameter signals the agent-layer protocol for this SVCB record (e.g. proto=a2a, proto=mcp). The companion proto-ver parameter carries the version of that protocol in the protocol's own syntax (e.g. proto-ver=1.0 with proto=a2a, proto-ver=2025-11-25 with proto=mcp). proto-ver is a selection hint only; the protocol's own negotiation settles compatibility. Whether version signalling in DNS provides useful signal or adds only RDATA bloat remains under evaluation pending interoperability testing.¶

5.2. Domain Control Validation

DNS-AID anticipates that agents and directories will sometimes need to prove that the publisher of an agent record controls the DNS zone in which it is published. The procedure in [I-D.draft-ietf-dnsop-domain-verification-techniques] is one candidate mechanism: the challenger issues a short-lived token that the claimant publishes at _agents-challenge.{domain} as TXT, and the challenger reads it back over a DNSSEC-validated path. A binding parameter such as bnd-req=svc:<agent>@<issuer> can scope a token to a specific agent leaf and issuer to prevent cross-challenger reuse.¶

A representative TXT record body and ABNF:¶

challenge   = token-attr 1*(SP attr)
token-attr  = "token=" 1*VCHAR
attr        = domain-attr / bnd-attr / expiry-attr
domain-attr = "domain=" 1*VCHAR
bnd-attr    = "bnd-req=svc:" 1*VCHAR "@" 1*VCHAR
expiry-attr = "expiry=" date-time   ; RFC 3339

_agents-challenge.acme.example. 60 IN TXT (
    "token=gjdqocfk4vhiq2bn" "domain=acme.example"
    "bnd-req=svc:api@acme.example" "expiry=2026-05-13T23:00:00Z" )

Open questions that should be answered before this is made normative: how a verifier scopes the number of records consulted for one verification, how expiry semantics interact with DNSSEC validity windows, and whether DCV results should be cached past their expiry. Consideration will also need to be given to the risk of this being used as an attack vector.¶

5.3. SVCB Indicating TLSA Support

An SvcParamKey in the agent SVCB record could be used to indicate TLSA support for the agent. As the SVCB query is the first one queried this might improve efficiency. Although the discovering agent may simply send both an SVCB and TLSA query simultaneously and infer support from the query responses, indicating the presence of a TLSA record would be more deterministic.¶

5.4. EDNS(0) Discovery Hints

A separate experimental mechanism allows a requestor to attach selector and metering hints to a DNS-AID query as an EDNS(0) option [RFC6891]. Substrate selectors (realm, transport, jurisdiction, minimum trust level) can influence the returned RRset and participate in cache keys; metering hints (intent class, freshness budget, parallelism) are advisory and do not fragment caches. A reference implementation and wire format are tracked in dns-aid-core (see docs/experimental/edns-signaling.md and the accompanying ABNF). No IANA option-code reservation is requested in this revision; the option uses the private-use code range.¶

5.5. Consolidated Registry and Cross-Domain Search

DNSAID-01 left open how a consolidated registry would operate -- whether it would scrape DNS-AID records published by individual organizations, or require attestation and verification from each publisher. The same question covers cross-domain search, i.e. locating agents whose origin domain the requestor does not know. This document treats both as upper-layer work; see [I-D.draft-narajala-courtney-ansv2] for one complementary proposal.¶

5.6. Operator and Service Metadata SvcParamKeys

Additional policy and realm SvcParamKeys are reserved for operator-controlled and service-controlled metadata that does not affect basic discovery: multi-tenant scoping, jurisdictional and compliance hints, and commercial signaling such as cost-per-unit-of-input, unit-of-input definitions, billing tiers, or rate-limit hints. The exact syntax of these payloads, the registry policy for any sub-keys that emerge, and the interaction between these parameters and the trust profile described in the Security Considerations section are open questions. Implementations are encouraged to experiment with concrete payloads in the private-use space and to report back so that a future revision can specify a common subset.¶

5.7. Connection Extensions as SvcParamKeys

Using connect-class and connect-meta provides a mechanism to signal which transport models to use, and with which pieces of metadata. Class may indicate fields like direct [RFC9460] for hostname+port semantics and lattice for integration with a vendor tools and services. Test implementations can use the Private Use parameter range and be moved to registered SvcParamKeys if they have wide spread applicability.¶

5.8. Zero Trust Extension SvcParamKeys

Adding a capability like enroll-uri may be relevant for mediated zero-trust overlay connections where a caller/service must first complete enrollment/authentication prior to invocation of any skills. This can be added via an SvcParamKey.¶

5.9. JSON-Encoded Organization Index

The TXT record-based form of _index._agents.{domain} may be suitable for small organizations and could be machine-parseable. A JSON-bodied variant, among others, has been discussed for larger inventories where TXT record chunking and multi-record assembly becomes complex and more subject to error. The encoding, signing, and validity-window rules for such a variant are open questions deferred to a future revision. It is anticipated that other standards will address organizational agent registries and the protocols to access them..¶

6. Security Considerations

This section discusses the security properties of DNS-AID itself; the security of the agent protocols in agent-to-agent communication is out of scope. The threat models in [RFC4033] (DNSSEC), [RFC9460], Section 10 (SVCB), and [RFC6698] (DANE) apply unchanged.¶

6.1. Authenticity, Integrity, and Trust

DNS-AID inherits authenticity from DNSSEC: a consumer that follows a validated chain of trust to a signed SVCB and TLSA record set has cryptographic evidence that those records were placed by an entity controlling the zone in which they were signed. DNS-AID does not, and cannot, assert that the agent reached through those records behaves benevolently. Authentic records may point at a malicious or compromised agent, and the descriptor referenced by the well-known or desc-ref parameters may itself contain instructions hostile to a consumer that interprets it (e.g. prompt-injection attacks against language-model agents). Consumers MUST treat the records published in DNS-AID as a verifiable transport for metadata, not as a trust signal in their own right; trust judgments MUST be made out of band by combining DNS-AID records with reputation, attestation, or organizational policy systems.¶

6.2. TLS Endpoint Authentication

DNS-AID consumers SHOULD authenticate the TLS endpoint advertised by an SVCB record using DANE TLSA records [RFC6698] [RFC7671] at the conventional _443._tcp.<owner> name. Three operational postures are common:¶

permissive (default for general-internet consumers): TLSA is queried; if a matching record is present and matches the presented certificate, the connection is pinned to it. If no TLSA record is present, the consumer falls back to WebPKI validation.¶
preferred: As permissive, but the absence of a TLSA record is noted in the consumer's audit log.¶
strict: A connection is refused unless a TLSA record is present and matches.¶

This document does not mandate any one posture. Strict-by-default is not appropriate for the general Internet today given current DANE adoption; permissive-with-fallback matches the implementation experience of [DNS-AID-CORE].¶

6.3. Downgrade Resistance

A publisher that requires a consumer to honor a particular SvcParamKey SHOULD list that key in the SVCB mandatory= parameter per [RFC9460], Section 8. A consumer that does not implement a key listed in mandatory= MUST skip the record. This prevents an attacker on the path from substituting an older subset of parameters without breaking discovery for old consumers. Until IANA assigns code points for the keys in this document, implementations may use different private-use numbers for the same key, so listing those keys in mandatory= (for example mandatory=proto) will not work reliably across implementations; TXT at the same owner name (Section 4) is the interoperability fallback in the meantime.¶

6.4. DNSSEC Dependency and Operational Pitfalls

DNS-AID is deployable without DNSSEC, but the authenticity guarantees it offers depend on a validated path from the root to the zone of interest. Consumers SHOULD validate DNSSEC and SHOULD refuse to act on bogus or unverifiable records. Zone operators SHOULD follow the operational guidance in [RFC6781] and SHOULD avoid signing algorithms or key sizes that resolvers in the target ecosystem are likely to ignore.¶

The mandatory= mechanism and DNSSEC validation interact: an unsigned record substituted in transit may cause a consumer to fall back to a less safe code path. Consumers MUST NOT relax DNSSEC validation requirements based on the apparent absence of mandatory= keys.¶

6.5. Threat-Model Cross-References

The threat catalogs maintained by the OWASP Multi-Agentic System Threat Modeling work [OWASP-MAESTRO] cover many agent-layer concerns that DNS-AID does not address directly. The substrate-level threats that DNS-AID does address (transport-identity spoofing, capability poisoning, downgrade) map roughly onto MAESTRO threats T47 / T7.1 / T9 (rogue server, agent impersonation), BV-2 (tool-description poisoning), and T7.6 (transport fallback downgrade). DNS-AID does not address workload-level MAESTRO threats; consumers SHOULD treat DNS-AID as one layer in defense in depth.¶

6.6. Privacy Considerations

The names queried during DNS-AID discovery are visible to recursive resolvers and to any on-path observer of unencrypted DNS traffic. Consumers SHOULD use DNS over TLS [RFC7858] or DNS over HTTPS [RFC8484] where available. EDNS Client Subnet [RFC7871] SHOULD NOT be sent on DNS-AID queries that would disclose user-identifying agent names; operators SHOULD set ECS scope to zero where DNS-AID is the only consumer of a zone.¶

Table 2: Underscored and Globally Scoped DNS Node Names
RR Type	_NODE NAME	Reference
SVCB	`_agents`	this document
SVCB	`_index`	this document
TXT	`_agents-challenge`	this document (experimental; see Section 5.2)

8. References

8.1. Normative References

[I-D.blwm-dawn-common-discovery-parameters]: Brockners, F., Jakab, L., Williams, N., and J. Mozley, "Common Agent DNS Discovery Parameters", Work in Progress, Internet-Draft, draft-blwm-dawn-common-discovery-parameters-00, 27 October 2026, <https://datatracker.ietf.org/doc/html/draft-blwm-dawn-common-discovery-parameters-00>.
[RFC2119]: Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, <https://www.rfc-editor.org/rfc/rfc2119>.
[RFC4033]: Arends, R., Austein, R., Larson, M., Massey, D., and S. Rose, "DNS Security Introduction and Requirements", RFC 4033, DOI 10.17487/RFC4033, March 2005, <https://www.rfc-editor.org/rfc/rfc4033>.
[RFC6698]: Hoffman, P. and J. Schlyter, "The DNS-Based Authentication of Named Entities (DANE) Transport Layer Security (TLS) Protocol: TLSA", RFC 6698, DOI 10.17487/RFC6698, August 2012, <https://www.rfc-editor.org/rfc/rfc6698>.
[RFC7671]: Dukhovni, V. and W. Hardaker, "The DNS-Based Authentication of Named Entities (DANE) Protocol: Updates and Operational Guidance", RFC 7671, DOI 10.17487/RFC7671, October 2015, <https://www.rfc-editor.org/rfc/rfc7671>.
[RFC8174]: Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.
[RFC8552]: Crocker, D., "Scoped Interpretation of DNS Resource Records through "Underscored" Naming of Attribute Leaves", BCP 222, RFC 8552, DOI 10.17487/RFC8552, March 2019, <https://www.rfc-editor.org/rfc/rfc8552>.
[RFC9364]: Hoffman, P., "DNS Security Extensions (DNSSEC)", BCP 237, RFC 9364, DOI 10.17487/RFC9364, February 2023, <https://www.rfc-editor.org/rfc/rfc9364>.
[RFC9460]: Schwartz, B., Bishop, M., and E. Nygren, "Service Binding and Parameter Specification via the DNS (SVCB and HTTPS Resource Records)", RFC 9460, DOI 10.17487/RFC9460, November 2023, <https://www.rfc-editor.org/rfc/rfc9460>.

8.2. Informative References

[ARD-SPECIFICATION]: "Agentic Resource Discovery Specification", n.d., <https://agenticresourcediscovery.org>.
[DNS-AID-CORE]: Infoblox, Inc. (contribution to Linux Foundation pending), "dns-aid-core: Reference implementation of DNS-AID", 2026, <https://github.com/infobloxopen/dns-aid-core>.
[I-D.draft-blwm-dawn-common-discovery-parameters]: Brockners, F., Jakab, L., Williams, N., and J. Mozley, "Common Agent DNS Discovery Parameters", Work in Progress, Internet-Draft, draft-blwm-dawn-common-discovery-parameters-00, 27 October 2026, <https://datatracker.ietf.org/doc/html/draft-blwm-dawn-common-discovery-parameters-00>.
[I-D.draft-ietf-dnsop-domain-verification-techniques]: Sahib, S. K., Huque, S., Wouters, P., Nygren, E., and T. Wicinski, "Domain Control Validation using DNS", Work in Progress, Internet-Draft, draft-ietf-dnsop-domain-verification-techniques-13, 21 June 2026, <https://datatracker.ietf.org/doc/html/draft-ietf-dnsop-domain-verification-techniques-13>.
[I-D.draft-jimenez-agent-directory]: Jimenez, J., "Agent Directory", Work in Progress, Internet-Draft, draft-jimenez-agent-directory-01, 8 May 2026, <https://datatracker.ietf.org/doc/html/draft-jimenez-agent-directory-01>.
[I-D.draft-mk-dnsop-svcb-well-known]: Mozley, J. and D. King, "An SVCB Service Parameter for Well-Known URI Paths", Work in Progress, Internet-Draft, draft-mk-dnsop-svcb-well-known-01, 19 October 2026, <https://datatracker.ietf.org/doc/html/draft-mk-dnsop-svcb-well-known-01>.
[I-D.draft-narajala-courtney-ansv2]: Courtney, S., Narajala, V. S., Huang, K., Habler, I., and A. Sheriff, "Agent Name Service v2 (ANS): A Domain-Anchored Identity Layer for Autonomous AI Agents", Work in Progress, Internet-Draft, draft-narajala-courtney-ansv2-02, 26 October 2026, <https://datatracker.ietf.org/doc/html/draft-narajala-courtney-ansv2-02>.
[OWASP-MAESTRO]: OWASP GenAI Security Project, "Multi-Agentic System Threat Modeling Guide v1.0", April 2025, <https://genai.owasp.org/resource/multi-agentic-system-threat-modeling-guide-v1-0/>.
[RFC6763]: Cheshire, S. and M. Krochmal, "DNS-Based Service Discovery", RFC 6763, DOI 10.17487/RFC6763, February 2013, <https://www.rfc-editor.org/rfc/rfc6763>.
[RFC6781]: Kolkman, O., Mekking, W., and R. Gieben, "DNSSEC Operational Practices, Version 2", RFC 6781, DOI 10.17487/RFC6781, December 2012, <https://www.rfc-editor.org/rfc/rfc6781>.
[RFC6891]: Damas, J., Graff, M., and P. Vixie, "Extension Mechanisms for DNS (EDNS(0))", STD 75, RFC 6891, DOI 10.17487/RFC6891, April 2013, <https://www.rfc-editor.org/rfc/rfc6891>.
[RFC7858]: Hu, Z., Zhu, L., Heidemann, J., Mankin, A., Wessels, D., and P. Hoffman, "Specification for DNS over Transport Layer Security (TLS)", RFC 7858, DOI 10.17487/RFC7858, May 2016, <https://www.rfc-editor.org/rfc/rfc7858>.
[RFC7871]: Contavalli, C., van der Gaast, W., Lawrence, D., and W. Kumari, "Client Subnet in DNS Queries", RFC 7871, DOI 10.17487/RFC7871, May 2016, <https://www.rfc-editor.org/rfc/rfc7871>.
[RFC8484]: Hoffman, P. and P. McManus, "DNS Queries over HTTPS (DoH)", RFC 8484, DOI 10.17487/RFC8484, October 2018, <https://www.rfc-editor.org/rfc/rfc8484>.