World ID Wants Into Banking. The Missing Primitive Is Revocation.

On October 18, World published a proposal to bring World ID into banking and financial services. Read it twice. There is no bank in it. No named partner, no pilot geography, no participant count, no implementation date, no third-party audit reference. An announcement aimed at regulated financial infrastructure that names zero regulated financial institutions. That absence is the load-bearing detail, and it is not the one World wanted discussed. The technical content is narrower and more interesting: World ID's three credentials — Proof of Human from the Orb's iris scan, Proof of Presence from selfie liveness, Proof of Attributes from government document attributes — are now formally decoupled. Each answers a different question. Uniqueness. Presence. Entitlement. Decoupling is correct engineering. It is also, inside a bank, insufficient in a way that the announcement quietly avoids. Logic does not bleed, but code leaves traces — and what the code traces here is a boundary, not a product. World's public arc has been a slow migration from consumer distribution to institutional plumbing. The Orb — a purpose-built biometric device — was always the unusual part of the design: Sybil resistance implemented in hardware rather than inferred from behavior or graph analysis. Most decentralized identity systems attempt uniqueness in software, and most of them fail, because software uniqueness is a social claim wearing a cryptographic costume. A physical scanner that binds one iris to one credential is a real answer to a real problem, and it is expensive in exactly the way that makes it credible. The October 18 framing extends that logic outward. Instead of asking individuals to prove they are human in order to receive tokens, World is asking institutions to accept a proof that a human is present and unique at the moment of an interaction. The adjacent narrative is AI agent payments — autonomous software initiating transactions on behalf of a principal, which creates a genuinely new authorization problem that this industry has not solved. The market backdrop matters. We are four months into a sideways tape, and chop is for positioning rather than conviction. Announcements in this regime get priced as optionality, not as fundamentals, which is precisely why the details of what was and was not shipped deserve a cold read. One sourcing caveat before the teardown: the document I am working from is commentary, signed to a news desk and an editor, with the publishing platform unmarked. Its NIST reference is an analogy drawn by the author, not a certification held by World. Treat the framing as opinion and the three credential definitions as the only hard facts. Start with what the credentials actually do, because the modularity is the most technically literate part of the design. Proof of Human asserts that a credential maps to one distinct living person. Proof of Presence asserts that a person is physically present at this moment rather than a recorded replay. Proof of Attributes asserts that the holder satisfies some property — a passport exists, an age threshold is met — without disclosing the underlying document. Three questions, three answers, composable, each minimizing disclosure relative to the alternative. Now count the questions a bank needs answered before it moves money. Is this a unique person. Is this person here. Does this person hold a verified attribute. And: is this specific person permitted to move this specific amount to this specific counterparty, right now, under a revocable mandate. That fourth question is not a credential. It is an authorization primitive, and the announcement does not ship one. The distinction is not semantic. The NIST digital identity framework treats biometrics as a single authentication factor inside a multi-factor construct, never as a standalone trust root. World's own materials, on close reading, appear to accept this: the credentials are positioned as inputs to a control system, not as the control system itself. If that reading is correct, the honest description of the product is a uniqueness factor embedded inside an existing KYC/AML stack — useful, defensible, and considerably smaller than the word 'banking.' Substitutability follows from that positioning, and it cuts both ways. If World ID is one factor among several, an institution can swap it for a competitor's uniqueness attestation without re-architecting its stack — low switching costs in both directions. Sticky infrastructure is what a trust root becomes when everything else depends on it. A factor stays a factor. The commercial ambition implied by 'banking' is closer to the former than the announcement supports. Here is where my own audit history sharpens the concern. In 2026 I worked a $50 million exploit on an AI trading agent platform. The failure was not identity. Every command in the exploit path came from an authenticated principal. The failure was scope: unverified model output was parsed into smart contract calls, and the permission model had no notion of a bounded mandate — no per-counterparty ceiling, no expiry, no independent kill path that did not require the agent's cooperation. Once the agent was compromised, the principal's authority was effectively infinite until a human noticed. That is the exact primitive World's banking pitch needs and does not visibly provide. For agent payments, the required set is unglamorous: an account-level scope, a spend ceiling, a counterparty allowlist, a time-bounded consent that does not silently renew, and — critically — a revocation channel that operates without the agent's assent. If revocation requires the agent to cooperate, it is not revocation. It is a request. The second structural problem is data flow. A zero-knowledge proof can minimize what is disclosed at the proof boundary. It cannot minimize what the surrounding service observes: account identifiers, counterparties, amounts, timing, device signals, and the fact of the query itself. The minimization is local, not systemic. For iris-derived templates specifically, the questions a compliance officer will ask are storage location, deletion path, and cross-border transfer — none of which the announcement answers. Volume is noise; the wallet cluster is signal, and here the signal is a proof boundary with an unexamined data lake behind it. Then there is the omission. A document framed around banking adoption and compliance boundaries says nothing about World's regulatory record — the inquiries and suspensions in Kenya, Spain, Portugal, Germany, and South Korea over biometric data handling. A compliance-facing document that omits the compliance history is not neutral. It is curated. No independent security audit is disclosed. No Orb hardware trust root documentation. No statement on whether iris templates are stored on device, in a cloud enclave, or deleted after derivation. Those are the three questions every prior regulatory inquiry converged on, and their absence from a document aimed at banks is a scoping decision, not an oversight. What is verified: the three credential types exist and run. What is unverified: any institutional deployment. The gap between those two statements is where the entire procurement question lives. The bulls are not wrong about the architecture. Decoupling uniqueness from presence from attributes is the correct primitive, and it is the opposite of how most identity projects ship — one monolithic verification that leaks everything in order to answer anything. Selective disclosure by construction is a genuine advance, and a hardware root of trust is currently the only mechanism that resists Sybil attacks at population scale without degrading into social graph inference. Credit also where the incentive structure is concerned. The conspicuous silence on token economics in a document about banking is itself a data point. A team seeking retail attention would have led with emission, APR, or airdrop mechanics. Leading with institutional compliance suggests value capture is migrating toward B2B verification fees — per-check pricing sold to banks, KYC vendors, and agent platforms. That is a slower curve with better retention than inflationary distribution, and it is a defensible position for a company whose consumer narrative has fatigued. The third thing the piece gets right, almost against its own promotional interest: it draws boundaries. In a cycle where every announcement inflates scope, a document that states 'identity proof is one component of a control system' does the reader a favor. Imagination is infinite, but liquidity is finite — and so is institutional patience for overstated capability. What a bank is actually buying here is not a credential. It is a liability question: who is accountable when an autonomous agent exceeds its mandate and the revocation path requires the agent's cooperation to execute. Gas fees are the price of truth, and the price of institutional adoption is a written answer to that question. The window is narrow. Agent payment infrastructure is being built now, and the revocation and spend-limit middleware layer is unoccupied. Whoever ships a standardized, agent-independent kill path in the next six to twelve months will own the default that banks eventually demand — and World, holding the uniqueness layer, is positioned to build it but has not yet said it will.