Безопасное подтверждение платежей AI-агента: четыре проверяемых рубежа
Description and use
The prototype shows how four security layers—bounded session, intent binding, authenticated single-use approval, and external signing—protect a payment AI agent from attacks involving substitution, replay, and bypassing approval checks.
Problem
Modal approval in payment AI agents does not prove identity, binding, freshness, or prevent credential misuse, leading to potential fraud.
Outcome
The prototype demonstrates that each security layer prevents specific attack vectors; removing any layer causes the corresponding test to fail.
Workflow
- 1. Create a bounded session with payee allowlist, currency, limits, expiry, and approver public key.
- 2. Agent requests payment; signer builds intent and stores it.
- 3. Signer computes intent_hash and signs a domain-separated message with approver’s private key.
- 4. Approver verifies signature, checks nonce, and ensures single-use.
- 5. Signer executes intent only after all checks pass, using external signing to access payment credentials.
- 6. Tests run with pytest to ensure each layer fails if removed.
Evidence
Secure human approval for a payment agent: bounded session, intent binding, authenticated one-time approval and external signing, each one with a test you can run.
The agent never gets general payment capability. It gets a session, a narrow and expiring grant created by a human or an admin process.
The signer builds a payment intent and stores it. The intent records payee ID, amount, currency, reference, session ID and a fresh intent ID.
The approval is an Ed25519 signature made with a private key that only exists on the approver's device.
- Record date: 2026-10-07T18:55:58.575Z