Review · November 21, 2026
Post-Quantum Requirements for Software and AI Agent Identity
What problem is this solving?
Agent delegation records get kept as audit evidence for years. The signatures on them are classical. Once a quantum computer can forge those signatures, every record that was never timestamped becomes deniable at the same moment.
The draft's central point is about records, not algorithms. A signature proves who signed, never when. The fix is the old archival practice of anchoring records in trusted time and renewing that protection. The catch is that an RFC 3161 timestamp is also a classical signature, so the anchoring has to become post-quantum too, by a stated date.
It also names a gap specific to OAuth. A verifier that fetches a JWKS over TLS has no record of what bound that key to a trust anchor. An auditor in 2040 can't tell which certificates protected a key in 2026.
How does this fit in with existing work?
This is a requirements document. It defines no protocol, no claims, and no IANA registrations. Its seven requirements (PQ-R1 to PQ-R7) are addressed to authors of other specs.
The WIMSE charter has no post-quantum work item. The closest fit is as input to the security considerations of AIMS, which names no signature algorithm at all. The audit half fits the proposed AUDIT working group better than WIMSE.
It has no normative dependencies beyond BCP 14. It leans informatively on Token Exchange, RFC 8414, the JWT BCP, identity chaining, RFC 4998 evidence records, and around twenty individual drafts.
One finding matters directly to OAuth. The draft says a chain built from Token Exchange alone does not conform. The act claim names who acted, not which algorithm signed, and nothing binds a new token to the tokens it was exchanged for.
What does it actually require?
- PQ-R1, agility. Explicit algorithm identifiers, and verifiers that accept several algorithms at once.
- PQ-R2, post-quantum issuance. Any credential type kept as evidence is signed post-quantum from a transition date "T" that each spec picks.
- PQ-R3, chain binding. Each hop is cryptographically bound to its parent. A token service does this by including a digest of each input token.
- PQ-R4, anchors first. Trust anchors migrate no later than the credentials under them, and verifiers stop accepting the classical anchor by T.
- PQ-R5, downgrade resistance. Verifiers reject algorithms outside local policy, even when the signature verifies.
- PQ-R6, chain visibility. The verifier sees the "algorithm profile" of every hop, including the algorithms binding each key to its anchor. A token service asserts the union of the profiles it accepted. After T, a token with no assertion is rejected.
- PQ-R7, record longevity. Every retained record is timestamped promptly, the timestamp becomes post-quantum by T, older records are back-anchored, and protection is renewed on a schedule.
So who has to change? Authorization servers would add a new assertion to every exchanged token and keep the input tokens. Resource servers would evaluate a set of algorithm profiles against policy. Anyone retaining records would run an RFC 4998 style archive with post-quantum timestamping authorities, which the draft admits are scarce today.
Will this lead to interoperable implementations?
Not by itself, and it doesn't claim to. The problem is that it leaves the interoperable parts undefined while placing MUSTs on them.
The algorithm profile has no representation. PQ-R6 depends on a token service asserting a set of profiles that a verifier in another organization can evaluate. The draft defines no claim name and no syntax. Appendix A writes entries like "ES256 / RSA-2048" informally. Two specs that each satisfy PQ-R6 will not understand each other's assertions.
Every spec picks its own transition date. The draft's own argument is that delegation crosses organizations with no shared policy. A chain crossing two domains with different T values has no defined behavior. The draft asks the working group whether T should be tied to policy dates, a fixed period, or left open. That question needs an answer before the fail-closed rules mean anything.
The JWKS gap is named but not closed. The draft lists x5c, OpenID Federation trust chains, and signed key sets as possible carriers, then declines to pick one. The fallback is that the retaining party saves its own copy of the JWKS and TLS chain, which the draft concedes is only good evidence for the party that saved it.
BCP 14 keywords are aimed at spec authors. "A specification MUST require that verifiers..." is an awkward construction in an Informational document. Nobody can test conformance to it.
What could be removed?
- PQ-R1 and PQ-R5. Section 7.1 admits these restate BCP 201 and the JWT BCP. A reference would do.
- Section 8.2. Three pages analyzing whether a component of a JOSE composite signature can verify as a standalone JWS. It ends with a question to the JOSE working group. That belongs on the JOSE mailing list, not here.
- Most of PQ-R7. The draft says Sections 4.6 and 4.7 of the agent audit record draft already satisfy PQ-R7(a) to (c), and that the renewal machinery belongs to RFC 4998. What's left is two additions: the anchor must be post-quantum, and by a stated date. That is a paragraph, not four pages.
- The policy timeline and draft survey. Sections 4.4 and 5 will be stale within a revision or two.
The parts worth keeping are short. Anchors migrate before leaves. A token service states which algorithms it accepted behind the token it issues. Retained records get post-quantum timestamps. The JWKS binding gets captured at issuance.
Writing quality
The reasoning is careful and the draft is honest about its limits. It credits prior art it missed in -00, narrows its own claims, and attributes specific constructions to other authors. It discloses that an LLM helped draft and search the literature.
It is far too long. 46 pages for seven requirements, nine of them references. The abstract is two dense paragraphs that restate the whole argument. The central argument appears in the abstract, the introduction, Section 3, and Section 6.2.
PQ-R7 is the hardest part to read. Clause (e) alone chains four conditions and two cross-references in one sentence. The summary table carries a disclaimer that the text governs where they differ, which tells you the text is hard to summarize. A reader can't easily extract a checklist from it.
The draft cites a large number of individual submissions as evidence of gaps. Most of those will expire. A requirements document should stand without them.
Who wrote it?
Venkata Karunakar Uppalapati (Independent Researcher) has no RFCs. This is their only Internet-Draft, active or expired. They are a new IETF author.
So what should the working group do with this?
Treat it as review input, not an adoption candidate. The observations about trust anchor ordering and unrecorded JWKS bindings are correct and worth a few sentences in the AIMS security considerations.
The Token Exchange finding deserves attention from OAuth. If a resource server needs to know which algorithms sat behind an exchanged token, that needs a registered claim with a defined syntax. This draft makes the case for one without writing it.
The retention requirements should go to the AUDIT discussion, where the audit record draft already carries most of them.