AIToday
Large Language ModelsAI Safety & AlignmentZenn AI/MLPublished: Nov 18, 2026, 04:00 JST

PAIP researcher asks if UCAN and OAuth RAR can cover personal AI delegation

Send on LINE
PAIP researcher asks if UCAN and OAuth RAR can cover personal AI delegation

3 Key Points

  1. What happened

    The researcher behind PAIP (Personal AI Interoperability Protocol) outlined six requirements — separate verification of an external action's authority and its legitimate basis, no promotion of untrusted text into commands, no skipping the receiving owner's permission, separating an 'I approved this' claim from independent proof, honoring revoked delegations, and returning the same ALLOW / DENY / UNKNOWN / UNSUPPORTED verdict from the same evidence.

  2. Why it matters

    There is no implementation or execution-based validation yet (zero), and the specification freeze has been postponed.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The scenario the researcher sketches is deliberately adversarial: Person A's AI is asked to send a message to Person B, gets steered by a third party's text, and claims 'A approved this' without A's actual approval. Person B's side can verify the signature and the delegation, but still has to decide separately whether the claim of approval is true and whether B's own owner permits the action. The harder problem is consistency — if B's AI and another company's AI receive the same request, they should return the same verdict, because 'deny' on one side and 'unsupported' on the other makes the results impossible to compare.

The six requirements are framed as a test for existing standards rather than a pitch for a new one. The researcher lists close candidates in two groups: delegation and revocation (UCAN, ZCAP-LD, Biscuit / Macaroons, GNAP, OAuth RAR) and decision and priority rules (XACML, Cedar, OPA), with approval bound to the operation (PSD2 dynamic linking, OpenID CIBA, FAPI) and separation of content from control (Dual LLM, CaMeL) as adjacent ideas. The working hypothesis is that no new protocol may be needed, and that the remaining question is an interoperability-profile problem: whether two independent implementations can combine existing technology with the same meaning. The researcher treats 'existing technology is enough' as a valid result.

The disclosed status is early. The specification has not been frozen, and there are no implementations. AI assistance was used to write and review the spec, with no independent third-party review yet; that AI-assisted review surfaced the case where authentication, expiry, unknown operations and missing evidence overlap and the verdict fails to resolve to one outcome — candidate fixes exist but have not been validated. The open call is for counterexamples and for arguments that a requirement itself is wrongly framed.

FAQ
Is PAIP a finished protocol?
No. The specification is at a prototype stage, the freeze has been postponed (BLOCK FREEZE), and there are zero implementations and zero execution-based validations.
What existing standards does the researcher compare against?
For delegation and revocation: UCAN, ZCAP-LD, Biscuit / Macaroons, GNAP (RFC 9635) and OAuth RAR (RFC 9396). For decision rules: XACML, Cedar and OPA.
What kind of feedback is being requested?
Pointed replies such as 'requirement Mxx is satisfiable with existing technology Y (basis: …)' or 'requirement Mxx breaks under condition Z (counterexample: …)', submitted to a GitHub issue.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAIエージェント guide: delegate email and research tasks, not whole jobs