CF-KYA v0.1 — Know Your Agent protocol

A protocol that answers one question a counterparty can check for itself: which legal person — a human or a corporation — stands behind this agent, for this action, right now, without trusting the agent's word.

Status — read this first

• Spec version: CF-KYA v0.1 (October 30, 2026).
• Implementation: in progress locally, NOT shipped, NOT deployed, NOT production-ready. Nothing in this dataset is a deployed service.
• Test vectors: structural vectors with expected verdicts. The signature fields in test_vectors.jsonl are clearly marked illustrative placeholders, NOT valid Ed25519 signatures. Cryptographically valid signed vectors ship with the verified v0.1 implementation, and only then. Do not use these vectors to validate a real verifier.
• Fictional data only: every principal, agent, issuer, and domain in the vectors is invented (*.example domains, reserved for documentation). No real person, company, key, or credential appears here.
• License: MIT.

The idea in one paragraph

An agent that acts for someone should be able to prove it, and a counterparty should be able to check the proof without phoning home to the agent's vendor. CF-KYA does that with signed credentials, delegation chains, and assurance levels — and deliberately nothing else: no new blockchain, no new token, no global registry. Key discovery copies DKIM: a principal publishes its public key at https://<domain>/.well-known/cf-kya.json. Verification needs only public keys, the credential, and a freshness-checked revocation list.

What is in this dataset

• CF_KYA_SPEC.md — the full CF-KYA v0.1 protocol specification: credential format, assurance levels L1/L2/L3, the five flows (register, present, verify, receipt, revoke), binding layers, delegation chains and swarm rules, interop notes, and the pre-registered ship bars.
• test_vectors.jsonl — 13 structural test vectors (valid, delegated child, forged, expired, revoked, wrong-scope, broken-chain scope, broken-chain expiry, L1-rejected, payload-replay, stale-revocation, swarm-valid, swarm-parent-revoked), each with an expected verdict and reason.
• build.py — the deterministic generator for the vectors, documenting the placeholder-signature policy.
• CRITIQUE.md — independent adversarial review (protocol, security, developer experience) with ranked holes and required fixes. Read it before adopting anything here; the "Known limitations" section below summarizes it.

What CF-KYA is not

• Not a KYC provider. It consumes attestations from existing identity providers and corporate registries. It stores evidence references, never identity documents.
• Not a fact-checker. It verifies who stands behind an action, not whether a claim about the world is true.
• Not a legal instrument by itself. The protocol supplies evidence; law supplies liability. The agency-agreement template behind a delegation needs a lawyer's blessing before real use, and no protocol creates legal liability on its own.
• Not a replacement for Visa's Trusted Agent Protocol, Mastercard's Verifiable Intent, Proof's x401, or SailPoint's enterprise agent identity. CF-KYA is the open, self-hostable version a small principal can adopt without joining a network or buying an enterprise suite. It is designed to sit alongside x401-style identity and x402-style payment, not to displace either.

Known limitations and open critique

This section is updated from independent review. The honest gaps, as of spec v0.1:

• Issuer trust is the trust anchor. A credential is only as good as its issuer's verification. A sloppy or malicious issuer can mint L3 credentials that verify mathematically. Verifiers must judge issuers, and the spec does not yet define an issuer reputation or audit mechanism.
• Key custody is the weak link. A stolen agent key is a stolen binding until expiry or revocation lands. Short credential life (24h max) bounds this; it does not eliminate it.
• Revocation freshness is a live dependency. Fail-closed verification refuses consequential actions when revocation data is stale — which is safe but means an unreachable revocation endpoint becomes a denial of service for legitimate agents. The freshness window is a policy choice the spec leaves to verifiers, and different choices will disagree.
• L1 is not identity. A self-asserted credential proves key control only. Any verifier that accepts L1 for a consequential action has misconfigured itself; the spec forbids it, but the spec cannot enforce it.
• Delegation depth and swarm accounting are bounded, not solved. The depth cap (4 links) and verifier-side narrowing checks bound the risk; a compromised parent still compromises its subtree until revocation propagates, and shared-budget accounting across a swarm needs an enforcement point (wallet/facilitator) outside this protocol.
• Adoption gaps found by independent developer-experience review (spec v0.1, not yet fixed): there is no live issuer or verifier, so a builder currently cannot complete a verifiable consequential call against a counterparty that checks it; there is no non-production test identity path, so a builder without a domain cannot start; the scope vocabulary (intent classes, tool names) is undefined, so two implementations cannot yet interoperate on scope; spend-cap semantics are undefined and no ledger design exists, so a global spend cap is currently unenforceable across counterparties; header names, the challenge flow, the canonical JSON algorithm, timestamp tolerance, and revocation TTL are not pinned in the spec; the relationship between the KYA credential and a separate governance token in one request is not specified; and the protocol offers no selective disclosure — display name and domain travel with every request. The single change that would most increase adoption is a live, self-serve test issuer plus a one-line fetch wrapper and verifier middleware, so a builder gets a verifiable test credential and a per-link verification in minutes. Until then, treat the adoption path as spec-only.

Origin

CF-KYA is the Bind layer of the Cognitive Firewall project (runtime governance for AI agent execution: verify claims, govern actions, bind actions to a principal). The Cognitive Firewall codebase itself is public at LunaTIKTOK/Cognitive-Firewall-Aperture on GitHub; this dataset holds the protocol specification and test vectors only, not that codebase.