At 07:18 on Monday, an operations lead is looking at a payments dashboard after a bad week. A compromised staff mailbox triggered a late Friday withdrawal request, a payment provider delayed settlement files, and the overnight reconciliation left a £42,000 difference. The money is not confirmed lost. The team still has to establish what moved, who approved it and whether the ledger can be trusted.
This is a composite scenario, not a report of a named firm. Its scale is typical of a mid-sized broker handling several thousand withdrawals a day. The security question is not simply whether the network was breached. It is whether one compromised identity or broken integration can move funds without detection, and whether the firm can restore a correct view of its obligations.
Start with the money path
Map each step from a client instruction to cash leaving a bank or payment provider. A common architecture has a client-facing balance service, a double-entry ledger, a payment orchestration service and one or more external accounts. These components must not be treated as interchangeable. The ledger records what the firm owes; bank and provider statements show cash movements. Neither view alone proves that the other is correct.
For each path, record the system that accepts the instruction, the identity that can authorise it, the service that submits it, and the evidence that confirms settlement. Include reversals, chargebacks, internal transfers, manual adjustments and provider callbacks. A callback that says a payment succeeded should be authenticated and checked against a known transaction, not accepted as a general instruction to increase a balance.
Threat modelling should cover both account takeover and ordinary operational mistakes. In this scenario, the weak point might be a mailbox, but the consequential permissions sit in the withdrawal service. Separate those failure domains. Practical questions include:
- Can a support agent change a bank beneficiary and approve a withdrawal in the same session?
- Can a service account submit arbitrary amounts, or only a payment tied to an authorised ledger entry?
- Can a retry create a second payment after a timeout, and does the system use an idempotency key?
- Can someone edit a ledger row directly, without an audited adjustment event and a second approval?
These questions are more useful than a list of security products. They expose where a single credential, message or human mistake could cross the boundary between an instruction and a movement of funds.
Make payment authority narrow
Withdrawal controls should use explicit states, such as requested, risk-checked, approved, submitted, settled and reconciled. Store each transition as an append-only event with the actor, timestamp, source and reason. A status change should not silently overwrite the earlier state. Where an external provider supports idempotency, use a stable key derived from the internal transaction ID and retain it across retries.
Set limits by channel, client risk and destination rather than relying on one global threshold. A new beneficiary, a changed contact detail or an unusual withdrawal pattern can trigger a hold or step-up check. A blanket delay on every payment can create service and regulatory problems, so use risk-based rules and document how staff can release a legitimate payment.
Dual approval is useful for high-risk actions, but it is not a cure if both approvers share a team account, device or compromised mailbox. The second approval should come from a separate named identity and be bound to the amount, beneficiary and transaction details. Any change to those fields should invalidate prior approval.
"A payment is not controlled because two people clicked approve. The approvals must bind to the same amount, destination and instruction."
Typical operating ranges vary by firm size, jurisdiction, payment rails and outsourcing model. The figures below are indicative market ranges for planning, not audited results or universal requirements.
| Control area | Typical market range | Operational meaning |
|---|---|---|
| Privileged human accounts covered by phishing-resistant MFA | 90% to 100% | Prioritise production, payment and identity administration access |
| High-risk withdrawals requiring independent approval | 100% above a defined threshold | Bind both approvals to the exact instruction and beneficiary |
| Payment and ledger exception alerting | 1 to 5 minutes | Alert on duplicate, delayed or unmatched events during operating hours |
| Ledger recovery point objective | 0 to 15 minutes | Use durable event capture and test replay from a known checkpoint |
| Critical payment-service recovery time objective | 1 to 4 hours | Set a tested target for restoring safe processing, not merely server uptime |
Protect identities, services and secrets
Use named accounts for people and separate identities for workloads. Enforce phishing-resistant MFA for administrators where available, short-lived privileged access, and just-in-time elevation for production changes. A small operations team may not be able to maintain a separate person for every control, but it can prevent one identity from changing a beneficiary, altering risk rules and releasing the resulting payment.
Service accounts deserve the same scrutiny. Restrict each credential to a specific API, environment and action. Keep secrets in a managed store, rotate them after exposure or role changes, and alert on use outside the expected workload. Avoid long-lived keys in build files, deployment scripts and developer laptops. Where a payment provider offers scoped credentials, use separate keys for production and testing, and ensure the test key cannot reach live accounts.
Network segmentation can limit impact, but it does not replace authorisation checks inside the service. A payment worker should verify a signed or otherwise trusted instruction from the ledger, check the current transaction state, and reject amount or beneficiary changes not represented in that instruction. Log security-relevant events to a separate system with restricted deletion rights. Do not put full payment credentials or unnecessary personal data in logs.
For teams reviewing their wider stack, Trading Technology's infrastructure coverage can help frame questions about vendors and platform boundaries. The control itself still needs an owner inside the firm. Outsourcing the payment rail does not outsource the firm's responsibility to understand how instructions, credentials and confirmations cross that boundary.
Reconcile continuously and rehearse recovery
Reconciliation is a security control as well as an accounting task. Match internal transaction IDs and amounts against provider acknowledgements, settlement files and bank statements. Treat missing, duplicated, out-of-order and altered records as separate exception types. A daily end-of-day check may be sufficient for low-volume channels, but a busy withdrawal service should detect material breaks during the day.
In the example, a £42,000 difference should produce a traceable exception, not a spreadsheet adjustment. The team needs to identify whether it came from a delayed settlement, a duplicate submission, a ledger defect or an unauthorised movement. Set thresholds that reflect normal timing and value, then page the responsible team when a break exceeds them. Keep the original records so a correction adds a new entry rather than erasing the evidence.
Recovery plans must restore integrity, not only availability. A clean failover that replays a payment event twice can turn an outage into a loss.
Use durable queues, idempotent consumers and tested replay procedures. Backups should be isolated from production credentials, protected against deletion, and restored on a schedule. Measure the time to resume safe payments and the amount of ledger activity that could need reconstruction.
During a suspected incident, the first actions should be controlled and recorded:
- Pause the affected payment route or impose a conservative withdrawal limit. Preserve unaffected channels if the evidence supports doing so.
- Revoke suspect sessions, tokens and credentials, then restrict beneficiary changes and manual ledger adjustments.
- Capture application, identity, provider and bank records before routine retention or log rotation removes them.
- Reconcile the affected window from the last trusted checkpoint, and require independent review before resuming normal processing.
- Assign one incident lead to coordinate engineering, operations, compliance and provider contacts, with a timestamped decision log.
Run this exercise with realistic failure conditions. A tabletop can test who has authority to stop payments; a controlled technical drill can test credential revocation, queue replay and restoration from backup.
Include the payment provider's escalation path and any bank cut-off times. Recovery targets that have never been tested are planning assumptions, not operational capabilities.
Set ownership before the next exception
Assign an owner to every control and define evidence that proves it works. That evidence might be a monthly report of privileged access, a record of sampled withdrawal approvals, reconciliation exception ageing, or results from a restore test. Track exceptions to closure. A policy that says withdrawals require two approvals offers little protection if the system cannot show which instruction each person approved.
Keep the control set proportionate. A firm with low payment volume may use a managed payment provider and a small number of manual checks, provided access is separated and daily reconciliation is reliable. A high-volume broker needs automated limits, event-level monitoring and tested failover because manual review cannot keep pace. The more complex the routing, entities and currencies, the more important it becomes to test each path separately.
The case for this advice is a firm that operates, directs or can materially alter the systems moving client money. It does not apply in the same way to a technology provider with no custody, payment authority or access to client funds; that provider should focus on its own data and service boundaries, while the broker or custodian owns the movement-of-funds controls.
Frequently Asked Questions
What is the most important security control for client funds?
No single control is sufficient. Separate payment initiation, beneficiary changes and approval, then reconcile authorised instructions against provider and bank records. This limits both unauthorised movement and undetected operational error.
How often should a broker reconcile client money movements?
Match payment events as they occur where volume and systems allow, and reconcile settlement files and bank statements at least daily. Set faster alerts for material exceptions, duplicates and missing confirmations.
Should withdrawal approvals always require two people?
Use independent approval for high-risk actions, with thresholds based on the firm's payment profile and obligations. Bind approval to the exact amount and beneficiary, and avoid shared identities or approval flows that one compromised account can control.
Platforms · Software · Providers · Markets