FinTechGenerator

Red Team Prompt to Test Your Wire Approval Path Against Cloned-Voice CEO Fraud

This Arbiter red-team prompt simulates the executive-impersonation payment chain against your live approval path in a controlled exercise, with findings routed to training and architecture, not discipline. Arbiter, the deepidv autonomous red agent, runs instruction injection with urgent, confidential payment requests from CEO and counterparty personas over consumer messaging channels, recording who treats the message as an instruction rather than a prompt to verify, then cloned-voice corroboration that answers callbacks to attacker-supplied numbers with synthetic voices, scoring whether callbacks route to directory numbers and whether scripts that demand specifics are followed. Ceremony pressure escalates urgency and seniority against identity-anchored approval, dual authorization, and action binding to show which stages hold mechanically and which bend to rank, a parallel beneficiary-change variant tests whether vendor banking-detail updates trigger re-verification from pre-existing contacts, and an evidence and recovery audit checks that each attempt's records would support same-day recall and insurer and law-enforcement engagement. It reports time to first unskippable verification per path, stages held versus bent, value at escape in the worst run, and the three cheapest fixes ranked by loss blocked per unit of treasury friction. Built for treasury, finance operations, and security leads who need to know whether their wire ceremony survives a convincing voice.

How to use this prompt

1
Open Arbiter in the deepidv dashboard and paste the full prompt, or run it in Claude, ChatGPT, or Gemini to model the attack chain before running the controlled exercise.
2
Add an INPUT section below the prompt describing each payment path, the authority-holders on it, the channels they accept requests on, your callback rules, and the verification stages between instruction and release.
3
Run the prompt and read time-from-instruction-to-first-unskippable-verification per path first: any path where a convincing message and voice reach release before a mechanical verification stage is the Fideuram failure mode.
4
Route findings to training and architecture, not discipline, as the prompt specifies, and hand the three ranked fixes to treasury with an owner for every stage that bent.
5
Re-run after any change to approval thresholds, callback procedures, or vendor banking-detail handling, and on a quarterly cadence alongside your executive exposure review.

The prompt

Arbiter, simulate the executive-impersonation payment chain against our live approval path in a controlled exercise, with findings to training and architecture, not discipline.

Attack sequence:
1. Instruction injection: deliver urgent, confidential payment requests to authority-holders over consumer channels, messaging personas of the CEO and of trusted counterparties, with insider-grade context, and record who treats the message as an instruction versus a verification prompt.
2. Cloned-voice corroboration: where callbacks occur, answer attacker-supplied numbers with synthetic voices of the impersonated parties, scoring whether callbacks route to directory numbers and whether specifics-demanding scripts are followed.
3. Ceremony pressure: escalate urgency and seniority against each verification stage, identity-anchored approval, dual authorization, action binding, measuring which stages hold mechanically and which bend to rank.
4. Beneficiary-change variant: run the quieter play in parallel, vendor banking-detail updates ahead of scheduled payments, testing whether changed details trigger new-beneficiary re-verification from pre-existing contacts.
5. Evidence and recovery audit: for every attempt, verify the ceremony's records (channel, callback, verification, approvers, bound details) would support same-day recall and insurer and law-enforcement engagement.

Report time-from-instruction-to-first-unskippable-verification per path, stages that held versus bent, value-at-escape in the worst simulated run, and the three cheapest fixes ranked by loss blocked per unit of treasury friction.

Test it in Claude or another LLM

This prompt is built for Arbiter inside deepidv, where Arbiter runs the impersonation chain against your live approval path in a controlled exercise. You can model the attack tree in any general LLM first with a synthetic payment path before running it against real people and systems.

1
Paste the full prompt into Claude, ChatGPT, or Gemini, replacing the opening 'Arbiter,' with a role instruction such as 'Act as a payment-fraud red-team lead modeling a cloned-voice executive impersonation chain.' Keep the attack sequence and the report line as written.
2
Add an INPUT section under the prompt and paste the synthetic sample block below so the model has payment paths, authority-holders, and ceremony stages to attack.
3
Add one framing line: 'This is synthetic test data. A callback to a number supplied in the request counts as no verification, since the attacker answers it.'
4
Check the output shape: time to first unskippable verification per path, stages held versus bent, value at escape in the worst run, and three fixes ranked by loss blocked per unit of treasury friction. If a stage is scored as holding without a mechanical control in the input, tighten the framing and re-run.
5
Once the shape is right, run it live in the deepidv dashboard where Arbiter executes the drill against your real approval path in a controlled exercise.

Synthetic sample data to paste alongside the prompt

Fake test data, safe to share with any LLM. Swap in your own once the output looks right.

PAYMENT PATHS (synthetic, fake):
- urgent outbound wire: chair can instruct finance directly, single approver below €20M (fake threshold), callback to the number in the request
- scheduled vendor payments: banking-detail changes accepted by email from the contact named on the change request
AUTHORITY-HOLDERS (fake): chair D. Example (reachable on WhatsApp), finance director E. Sample (email and desk phone)
VERIFICATION STAGES (fake): dual authorization above €20M only, no identity-anchored approval, no action binding
RECOVERY (fake): no pre-written bank recall runbook
OPEN ITEM (fake): whether a senior executive can waive the callback under time pressure

Pairs with on deepidv

Sources & further reading

FAQ

What is cloned-authority wire fraud?

It is executive impersonation aimed at payment authority: a spoofed message from a CEO or trusted counterparty requests an urgent, confidential transfer, and a cloned voice corroborates it when someone calls back on the attacker's number. Nothing technical is compromised, so the attack succeeds or fails entirely on whether the verification ceremony holds against urgency and rank.

What does the Cloned Authority Wire Drill measure?

Time from instruction to the first verification stage that cannot be skipped, per payment path; which stages held and which bent under pressure; value at escape in the worst simulated run; and whether the records would support same-day recall. It closes with the three cheapest fixes ranked by loss blocked per unit of treasury friction.

Why run the beneficiary-change variant in parallel?

Because it is the quieter play: instead of a dramatic urgent wire, the attacker updates a vendor's banking details ahead of a scheduled payment and lets the normal payment run do the work. The drill tests whether changed details trigger new-beneficiary re-verification through pre-existing contacts, not through the contact details supplied with the change.

Run it with live verification data

These prompts work in any LLM. Inside the deepidv dashboard, Luna, Arbiter, and Arc run them against your real sessions, screening lists, and audit trails.

Book a Demo