The Scope Creep Automated reviews of new AI-related working group submissions

Review · August 21, 2026

Privacy Considerations for DAWN

draft-iannone-dawn-privacy-considerations -00View on Datatracker →

Problem Statement

The core insight this draft is trying to document: the act of discovery itself leaks sensitive information. When an AI agent queries a discovery substrate for capabilities or resources, that query can reveal the querying entity's intentions — a medical AI searching oncology datasets reveals health context, a company agent searching GPU clusters before a product launch reveals business strategy. This is distinct from protecting the published properties of discoverable entities or the content of communications between them.

How This Fits Into the DAWN Landscape

This is an individual submission informational draft, not a protocol specification. It sits alongside the DAWN problem statement and terminology drafts as a companion considerations document, intended to inform the design of whatever discovery protocol(s) eventually emerge from the DAWN working group.

It does not extend any existing RFC, define any mechanisms, or establish new roles. Instead, it applies the RFC 6973 privacy threat analysis framework to the DAWN discovery context and introduces Private Information Retrieval (PIR) as a candidate mitigation technology.

The DAWN problem statement and terminology drafts are listed as normative references, which is somewhat unusual for an informational analysis document — normative dependencies imply this draft cannot be fully interpreted without those drafts, which may be appropriate but worth reviewing as those documents evolve.

High-Level Summary

The draft applies the RFC 6973 structured threat taxonomy to AI agent discovery:

  • Surveillance: Discovery infrastructure becomes a surveillance point — an observer watching query traffic can build a picture of an organization's interests and operational patterns.
  • Stored Data Compromise: Query logs and metadata retained by discovery systems become targets; minimization is recommended.
  • Correlation: Repeated queries and stable agent identifiers can be linked to profile entities or their operators over time.
  • Identification: Discovery protocols can inadvertently expose organization names, infrastructure ownership, or capability fingerprints beyond what is operationally necessary.

The draft also notes that the severity of these concerns is context-dependent: privacy requirements may be relaxed in closed, single-vendor ecosystems (internal coding assistants, HR workflow agents), while remaining critical for systems that interact with external parties.

As a mitigation technology, the draft introduces Private Information Retrieval (PIR) — a cryptographic technique allowing a client to retrieve an item from a server's database without the server learning which item was retrieved. The draft suggests PIR could prevent discovery registrars from learning what a querying agent is looking for, while acknowledging that scaling PIR to the size DAWN envisions needs further investigation.

Completeness and Usefulness as an Input Document

This is a very early skeleton draft. Roughly half of the planned sections are marked TBD, including:

  • Intrusion (Section 5.1.3)
  • Misattribution (Section 5.1.4)
  • Secondary Use (Section 5.2.3)
  • Disclosure (Section 5.2.4)
  • Exclusion (Section 5.2.5)
  • DNS privacy similarities (Section 6)
  • Privacy vs. Auditability (Section 7)
  • RFC 6973 compliance questionnaire (Section 8)
  • Security Considerations (Section 11)

The draft acknowledges that DAWN-specific mitigation categories will be added in future revisions. In its current state, it is more a statement of intent and a placeholder outline than a substantive analysis. The sections that are written (Surveillance, Stored Data Compromise, Correlation, Identification) provide reasonable framing but are brief — a paragraph each.

The RFC 2119 requirements language section appears unnecessary for an informational threat analysis document. No normative requirements are stated in the document as written, and the entire purpose of the document is descriptive analysis rather than protocol specification. This section should either be removed or justified once actual requirements language is added.

Writing Quality

The draft has several grammatical issues:

  • "It provides an threat analysis" (Section 1) — should be "a threat analysis"
  • "privacy considerations about published information will discussed" (Section 1) — missing "be"
  • "goes through the the privacy threats" (Section 5) — duplicated word
  • "linked to identify users or entities behavior" (Section 5.2.1) — awkward phrasing
  • "their use in DAWN to prevent registrars or entities hosting capacities of interest to retrieve information" (Section 9.1) — missing verb, sentence is incomplete

The draft is appropriately short for what it currently contains, but several sections are essentially empty placeholders. The introduction is well-motivated and the examples (medical AI querying oncology datasets, company agent searching GPU clusters) are effective and concrete. The PIR discussion is brief but interesting; it would benefit from even one sentence explaining at a high level how PIR works before diving into deployment considerations.

Author Background

Luigi Iannone ([email protected]) — Huawei Technologies France. 7 published RFCs. Currently chairs both the LISP and SIDROPS working groups and serves as an OPSDIR reviewer. An experienced and active IETF participant. Not a new author.

Antoine Fressancourt ([email protected]) — Huawei Technologies France. 0 published RFCs. This draft is his only active Internet-Draft. He serves as a reviewer in the Internet Area Directorate (intdir). Effectively a new IETF author making his first contribution.

Overall Assessment

The problem being identified — that query privacy in AI agent discovery is a distinct and important concern not covered by transport-layer security alone — is a valid and underexplored area. The RFC 6973 framework is the right tool for this analysis. However, with approximately half the document marked TBD and several grammatical errors, this draft is not ready for substantive review or adoption. It would benefit significantly from completion before being presented to the working group. The PIR suggestion is intriguing but speculative; it would benefit from a comparison to existing DNS privacy techniques (DoH, DoT, ODNS) to ground the discussion in prior IETF work, which the draft promises in Section 6 but has not yet delivered.