How do I verify a human authorized an AI agent action?
Verify it with a per-action cryptographic signature, not a login-time identity check. When an AI agent reaches a sensitive action, Yuthent requires a fresh, hardware-bound signature from the specific human the action is attributed to, produced on that human's own device and bound to the exact action. Your backend verifies the signature before the action runs. If no valid signature is present, the action does not execute.
The distinction that matters
A policy gate records that an action was allowed. A login check records that someone authenticated earlier. Neither proves who approved the action that moved the money, prescribed the dose, or granted the privilege.
Yuthent produces that proof: a hardware-bound, action-bound signature of a specific human, generated at the moment of execution, that an attacker cannot forge and an agent cannot fabricate.
How the verification works
- 01The agent reaches a sensitive action. Your backend policy classifies it as one that requires a human signature.
- 02A request goes to the specific human's enrolled device. The human approves with a fresh biometric, handled by the operating system inside trusted hardware.
- 03The device signs a description of the exact action with a hardware-bound key. The biometric never leaves the device. Only the signature and device-state attestation are sent.
- 04Your backend verifies the signature against a published public key, checks it is bound to this action, and confirms it has not been replayed. Only then does the action execute.
The signature is bound to the action payload, so a mutated parameter after approval breaks the proof. The result is evidence a third party can verify after the fact, offline, against a published public key.
Why regulated and public-sector buyers ask this
EU AI Act Article 14 requires meaningful human oversight of high-risk AI systems. A logged click or an approval dialog is oversight on paper. A per-action signature from a verified human is oversight that holds up as evidence. For payments, DORA and PSD2 SCA dynamic linking point in the same direction: the act, not the session, is what must be provable.
Yuthent is Execution Authority Infrastructure. It runs alongside your existing identity and agent stack, and every engagement starts with a scoped first deployment.
Common questions
How do I verify that a human authorized an AI agent action?
Require authority the named human signed, produced on that human's own device: either a signature over the exact action payload, or a standing mandate whose signed bounds the action falls inside. The backend verifies against a published public key before the action executes. A login-time identity check cannot answer this, because it proves who started a session, not who approved this action now.
Why is a login-time biometric not enough for agent actions?
A login-time biometric proves a human authenticated once, at the start. The authority then inherits to every downstream action, including actions an autonomous agent takes later inside that session. Yuthent moves the proof to execution time, so each sensitive action carries its own fresh, hardware-bound human signature.
What happens when an agent acts without a valid signature?
The action fails closed. If no valid per-action signature is present for an action that requires one, the backend does not execute it. A prompt-injected or compromised agent cannot fabricate the signature, because the signing key is held in device hardware and gated by a fresh biometric the agent does not control.