iProov’s experimental HAPS specification aims to verify human approval before AI agents execute sensitive or high-risk actions.
Biometric identity verification provider iProov has published the Human Approval and Presence Specification (HAPS), an experimental framework designed to help organizations verify that a human has approved a specific action before an autonomous AI agent can execute it.
Released on GitHub under the Apache-2.0 license, HAPS provides rules for linking verifiable human approval to a specific action requested by an AI agent. iProov has also released a partial Rust reference implementation and test vectors, inviting developers and industry experts to review, test and independently implement the specification.
The framework addresses an emerging challenge in agentic AI: an AI agent may have legitimate access to tools, services or delegated credentials while still taking an action that the human did not actually intend to approve. Potential causes include prompt injection, excessive goal-seeking or misuse of delegated permissions.
Verifying human approval
HAPS is designed for situations where organizations determine that an AI-generated action is sensitive enough to require additional human authorization. Rather than requiring approval for every agent action, the framework focuses on selected high-risk or critical operations.
When approval is required, the proposed process is to pause the action, present the intended action to the human, obtain evidence of genuine human presence and approval, and verify that the approval corresponds to the specific action before allowing it to proceed.
The specification is proof-agnostic, meaning it does not mandate a particular technology for proving human presence. iProov said biometric liveness is one possible implementation.
The company is releasing HAPS as an experimental specification and is encouraging the technology community to scrutinize the design, provide feedback and develop independent implementations.
As AI agents move from generating information to independently performing tasks, frameworks such as HAPS highlight the growing need to distinguish between an agent being authorized to act and a human explicitly approving a particular action.
