“A human approved it” is usually a claim, not evidence. On 22 October, iProov shipped a service that turns it into a signed credential bound to one specific action an agent was trying to take.
What happened
iProov launched a hosted human approval service for AI agent gateways, the first implementation of the Human Approval and Presence Specification (HAPS), which the company published as an open specification in September 2026. The launch integrations are agentgateway, the open-source agent gateway from solo.io, and Cerbos, an authorization policy engine, as the first policy-engine partner.
The flow, as described in the release:
- A policy engine decides that a given agent action needs human approval, and holds the action.
- iProov’s hosted service shows the named approver the exact action and its scope.
- If the person approves, they complete an iProov genuine human presence check (liveness, not a password or a click).
- iProov issues a signed, single-use credential bound to that exact action, and only then is the agent’s action released.
- If any detail changes — the release gives the example of a payment’s amount or destination — the credential no longer matches and the action is blocked.
The company frames the point bluntly: the credential is “verifiable proof that the right person approved this action, which no agent is capable of producing on its own behalf.” Andrew Bud, iProov’s founder and CEO, puts the same idea as “no matter how smart they become, agents will never have faces.”
The specification, a partial Rust reference implementation and test vectors are published at github.com/iProov/HAPS (Apache-2.0). The company also cites Microsoft’s 19 October Security blog, which called for policies that can “require cryptographically provable human approval for a high-impact step” — the same gap, described from the platform side.
Why it matters for KYA
Know Your Agent has two halves, and the industry has been building only one. Most agent-identity work answers who is this agent — SPIFFE IDs, X.509 certificates, workload attestation, registries. Far less answers who authorised this particular action, and can you prove it later. A gateway that can verify an agent’s identity but only takes a human’s word for approval is one prompt injection away from a defensible-looking incident.
Binding a credential to an action is the part worth copying. Approval that is scoped to an intent hash, audience-restricted, expiring and single-use is a mandate at action granularity rather than a session-wide permission. It closes the gap where consent to “move some money” quietly authorises a different amount to a different account. That is the difference between a policy and a receipt.
Liveness is aimed at the right adversary. The approval step is now the soft target: an agent with access to the user’s device, session and push notifications can satisfy an OTP or an in-app “Approve” button by itself. The HAPS repository treats this as a first-class property — it separates agent-resistant factors (hardware keys, biometric passkeys, liveness) from agent-accessible ones, and warns that “the method name alone does not establish that resistance.”
It is deliberately factor-agnostic. HAPS composes with WebAuthn, passkeys, smart cards and OAuth rather than replacing them, and the spec says a relying party can demand evidence of agent resistance independently of the assurance level a provider claims. That leaves room for the control to outlive any single vendor — which is the only way this becomes infrastructure rather than a product feature.
The spec is unusually honest about what it does not prove. Its own README states that it is an experimental draft and not production ready; that a matching `presentation_hash` “does not establish what appeared on screen or what a person perceived or approved”; that replay protection depends on the verifier’s policy and atomic consumption state; and, most importantly, that “a presence assertion does not establish authorization” — standing authority and scope belong to a mandate or capability layer above it. Read that as the correct division of labour: HAPS evidences fresh human approval of a described action, and something else must still say the agent was allowed to request it.
What a practitioner should take from it
- Decide which actions require approval at all. High-consequence, hard-to-reverse, cross-party: yes. Everything: the approval fatigue will defeat the control, and agents will learn to route around it.
- Require agent-resistant factors for anything that matters. A push notification to a device the agent can read is not human oversight.
- Verify the credential, do not log the claim. Check audience, expiry and single use, and store the signed credential next to the agent’s identity and the delegation it acted under — that trio is what answers “who did this and who approved it” six months later.
- Keep the mandate layer separate. Approval proves a person said yes to one action; it does not grant scope, and it does not tell you the agent was the agent it claimed to be.
Caveats: this is a vendor-hosted service with two launch integrations and no published independent security review; the specification is weeks old and explicitly experimental, and its own authors caution that passing the supplied tests does not establish conformance or that a human saw the right thing. The action description shown to the approver is the trust anchor, and it is supplied by the system asking for approval.
Sources
- iProov, “iProov Launches New Human Authorization Service for AI Agent Gateways with First Integrations”, press release via Business Wire, 22 October 2026 — businesswire syndicated copy.
- iProov, Human Approval and Presence Specification (HAPS) — draft specification, schemas, test vectors, `haps-v0.4.0`, factor registry and limitations — github.com/iProov/HAPS.
- iProov, HAPS reference implementation — github.com/iProov/HAPS-reference.
- Microsoft Security Blog, “Agents are your newest insiders: security must move at their speed”, 19 October 2026 — microsoft.com.
Note: the mechanism above is described from iProov’s release and the HAPS draft; no independent evaluation of the liveness check or the credential format has been published.
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.