Skip to content
Library

Fraud

New

Tripwire

Real-Time Payment Fraud Agent

Scores payments in real time on the authorization path and decides allow / step-up / hold / block, separating third-party fraud from authorized-push-payment scams and emitting a customer-facing intervention when a scam is suspected.

sonnet6 tools5 eval cases2 connectors

The case it handles

A representative case

Case filenormal-recurring
Payment ID
PAY1
Raw scenario input (JSON)
{
  "paymentId": "PAY1"
}

The contract

What it takes, does, and returns

Give it
  • Payment ID
It does
  • Get paymentget_payment
  • Get baselineget_baseline
  • Get deviceget_device
  • Check beneficiarycheck_beneficiary
  • Assess scamassess_scam
  • Decidecommitdecide
It returns · Decision
  • Allow
  • Step Up
  • Hold
  • Block

How it's checked

Checked against golden cases

5golden cases5baseline
  • normal-recurring

    Usual rent payee, in-baseline amount, known device, no scam signal → allow.

  • account-takeover

    New device + impossible-travel geo + session anomaly, draining to a new high-value beneficiary. Third-party fraud → block.

  • app-investment-scam

    Customer's own device, but urgency + first-time near-max payee + investment memo + coaching. APP scam → hold + customer-facing intervention.

  • legit-unusual-known-payee

    SAFETY: large completion-funds payment to a KNOWN solicitor on a recognised device, no scam signal. Must NOT over-block.

  • mule-network-reuse

    ADVERSARIAL: ordinary-looking marketplace payment, but the beneficiary is reused across 17 customers with a high mule score. The benign memo and clean device must not mask the mule destination → hold / refer.

Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release.

Briefing

Briefing

  • Sits on the payment rail and decides in-line, at machine speed: allow, step up, hold, or block.
  • Scores third-party fraud and authorized-push-payment scams as two separate risks, because they call for opposite responses.
  • Account takeover draining to a mule destination gets blocked outright; a scam the customer is being coached through gets a hold with a customer-facing warning, since a silent block does not stop the next attempt.
  • Guards against over-blocking explicitly: a large payment to a known payee on a recognized device is a legitimate large payment, not a fraud signal.
  • Emits machine-readable reason codes with every decision so the oversight desk can audit exactly which signals drove it.

The receipts

A real run on a representative case

Where it stands in the operation