The experimental specification helps organizations verify that a person approved a specific high-risk action before an AI agent carries it out.
iProov has published the Human Approval and Presence Specification, or HAPS, an experimental procedural framework for organizations that want stronger human oversight of AI agents.
The specification addresses a specific governance problem: an AI agent may have permission to access a tool or service, but that does not necessarily mean a human intended or approved the specific action the agent is about to perform. HAPS is intended to give organizations a way to verify that approval before selected actions proceed.
Published on GitHub under the Apache-2.0 license, HAPS includes procedural rules, test vectors, and a partial Rust reference implementation. iProov invites security researchers, developers, and other industry participants to review the work, challenge its assumptions, and explore independent implementations.
Permission is not the same as approval
The framework focuses on cases where an AI agent remains within its authorized permissions but still takes an unintended action.
The release cites risks such as prompt injection, excessive goal-seeking, and misuse of delegated credentials. In those situations, the system receiving the request may be able to verify that the agent is technically allowed to act, but still lack evidence that a human actually approved that specific action.
“AI agents are moving rapidly from answering questions to taking actions on our behalf. As their autonomy and capabilities grow, governance must keep pace,” said Andrew Bud, founder and CEO of iProov.
“We need to distinguish between an agent having permission to act and a human actually approving the specific action it is about to take.”
Human approval only where it matters
HAPS is not intended to require a human sign-off for every action an AI agent performs.
Instead, organizations decide which actions are sensitive or critical enough to require additional approval and what level of evidence should be required. This preserves the benefits of automation while avoiding constant prompts that could create approval fatigue.
When an additional approval step is triggered, the process is straightforward: pause the action, show the user exactly what the agent wants to do, obtain evidence of genuine human presence and approval, and verify that the approval matches the requested action.
The approval is then securely linked to that specific action, allowing the relying system to confirm that the user authorized what the agent is actually attempting to execute.
Proof method remains flexible
HAPS is proof-agnostic, meaning it does not mandate one particular method for proving human presence.
iProov has developed an internal implementation using biometric liveness as one possible example, but the specification itself is intended to support other forms of proof.
That design makes HAPS more of a governance and verification protocol than a single authentication technology.
Open for industry review
HAPS is being released at an experimental stage rather than as a finished industry standard.
Bud said the decision to publish it openly is intended to encourage scrutiny and collaboration as organizations work through the governance challenges created by increasingly autonomous AI agents.
The HAPS specification, reference implementation, and test materials are available through the iProov HAPS GitHub repository.